Line | Branch | Exec | Source |
---|---|---|---|
1 | /* | ||
2 | * HEVC video decoder | ||
3 | * | ||
4 | * Copyright (C) 2012 - 2013 Guillaume Martres | ||
5 | * | ||
6 | * This file is part of FFmpeg. | ||
7 | * | ||
8 | * FFmpeg is free software; you can redistribute it and/or | ||
9 | * modify it under the terms of the GNU Lesser General Public | ||
10 | * License as published by the Free Software Foundation; either | ||
11 | * version 2.1 of the License, or (at your option) any later version. | ||
12 | * | ||
13 | * FFmpeg is distributed in the hope that it will be useful, | ||
14 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | ||
15 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU | ||
16 | * Lesser General Public License for more details. | ||
17 | * | ||
18 | * You should have received a copy of the GNU Lesser General Public | ||
19 | * License along with FFmpeg; if not, write to the Free Software | ||
20 | * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA | ||
21 | */ | ||
22 | |||
23 | #include "get_bits.h" | ||
24 | #include "hevcdec.h" | ||
25 | |||
26 | #include "bit_depth_template.c" | ||
27 | #include "hevcdsp.h" | ||
28 | |||
29 | 74598 | static void FUNC(put_pcm)(uint8_t *_dst, ptrdiff_t stride, int width, int height, | |
30 | GetBitContext *gb, int pcm_bit_depth) | ||
31 | { | ||
32 | int x, y; | ||
33 | 74598 | pixel *dst = (pixel *)_dst; | |
34 | |||
35 | 74598 | stride /= sizeof(pixel); | |
36 | |||
37 |
2/2✓ Branch 0 taken 327424 times.
✓ Branch 1 taken 37299 times.
|
729446 | for (y = 0; y < height; y++) { |
38 |
2/2✓ Branch 0 taken 4382080 times.
✓ Branch 1 taken 327424 times.
|
9419008 | for (x = 0; x < width; x++) |
39 | 8764160 | dst[x] = get_bits(gb, pcm_bit_depth) << (BIT_DEPTH - pcm_bit_depth); | |
40 | 654848 | dst += stride; | |
41 | } | ||
42 | 74598 | } | |
43 | |||
44 | 27980362 | static av_always_inline void FUNC(add_residual)(uint8_t *_dst, const int16_t *res, | |
45 | ptrdiff_t stride, int size) | ||
46 | { | ||
47 | int x, y; | ||
48 | 27980362 | pixel *dst = (pixel *)_dst; | |
49 | |||
50 | 27980362 | stride /= sizeof(pixel); | |
51 | |||
52 |
2/2✓ Branch 0 taken 106549588 times.
✓ Branch 1 taken 13990181 times.
|
241079538 | for (y = 0; y < size; y++) { |
53 |
2/2✓ Branch 0 taken 1377565424 times.
✓ Branch 1 taken 106549588 times.
|
2968230024 | for (x = 0; x < size; x++) { |
54 | 2755130848 | dst[x] = av_clip_pixel(dst[x] + *res); | |
55 | 2755130848 | res++; | |
56 | } | ||
57 | 213099176 | dst += stride; | |
58 | } | ||
59 | 27980362 | } | |
60 | |||
61 | 16228982 | static void FUNC(add_residual4x4)(uint8_t *_dst, const int16_t *res, | |
62 | ptrdiff_t stride) | ||
63 | { | ||
64 | 16228982 | FUNC(add_residual)(_dst, res, stride, 4); | |
65 | 16228982 | } | |
66 | |||
67 | 7288426 | static void FUNC(add_residual8x8)(uint8_t *_dst, const int16_t *res, | |
68 | ptrdiff_t stride) | ||
69 | { | ||
70 | 7288426 | FUNC(add_residual)(_dst, res, stride, 8); | |
71 | 7288426 | } | |
72 | |||
73 | 3308668 | static void FUNC(add_residual16x16)(uint8_t *_dst, const int16_t *res, | |
74 | ptrdiff_t stride) | ||
75 | { | ||
76 | 3308668 | FUNC(add_residual)(_dst, res, stride, 16); | |
77 | 3308668 | } | |
78 | |||
79 | 1154286 | static void FUNC(add_residual32x32)(uint8_t *_dst, const int16_t *res, | |
80 | ptrdiff_t stride) | ||
81 | { | ||
82 | 1154286 | FUNC(add_residual)(_dst, res, stride, 32); | |
83 | 1154286 | } | |
84 | |||
85 | 54126 | static void FUNC(transform_rdpcm)(int16_t *_coeffs, int16_t log2_size, int mode) | |
86 | { | ||
87 | 54126 | int16_t *coeffs = (int16_t *) _coeffs; | |
88 | int x, y; | ||
89 | 54126 | int size = 1 << log2_size; | |
90 | |||
91 |
2/2✓ Branch 0 taken 9989 times.
✓ Branch 1 taken 17074 times.
|
54126 | if (mode) { |
92 | 19978 | coeffs += size; | |
93 |
2/2✓ Branch 0 taken 56315 times.
✓ Branch 1 taken 9989 times.
|
132608 | for (y = 0; y < size - 1; y++) { |
94 |
2/2✓ Branch 0 taken 523168 times.
✓ Branch 1 taken 56315 times.
|
1158966 | for (x = 0; x < size; x++) |
95 | 1046336 | coeffs[x] += coeffs[x - size]; | |
96 | 112630 | coeffs += size; | |
97 | } | ||
98 | } else { | ||
99 |
2/2✓ Branch 0 taken 124352 times.
✓ Branch 1 taken 17074 times.
|
282852 | for (y = 0; y < size; y++) { |
100 |
2/2✓ Branch 0 taken 1049312 times.
✓ Branch 1 taken 124352 times.
|
2347328 | for (x = 1; x < size; x++) |
101 | 2098624 | coeffs[x] += coeffs[x - 1]; | |
102 | 248704 | coeffs += size; | |
103 | } | ||
104 | } | ||
105 | 54126 | } | |
106 | |||
107 | 864282 | static void FUNC(dequant)(int16_t *coeffs, int16_t log2_size) | |
108 | { | ||
109 | 864282 | int shift = 15 - BIT_DEPTH - log2_size; | |
110 | int x, y; | ||
111 | 864282 | int size = 1 << log2_size; | |
112 | |||
113 |
2/2✓ Branch 0 taken 422855 times.
✓ Branch 1 taken 9286 times.
|
864282 | if (shift > 0) { |
114 | 845710 | int offset = 1 << (shift - 1); | |
115 |
2/2✓ Branch 0 taken 1765132 times.
✓ Branch 1 taken 422855 times.
|
4375974 | for (y = 0; y < size; y++) { |
116 |
2/2✓ Branch 0 taken 8083472 times.
✓ Branch 1 taken 1765132 times.
|
19697208 | for (x = 0; x < size; x++) { |
117 | 16166944 | *coeffs = (*coeffs + offset) >> shift; | |
118 | 16166944 | coeffs++; | |
119 | } | ||
120 | } | ||
121 | } else { | ||
122 |
2/2✓ Branch 0 taken 74992 times.
✓ Branch 1 taken 9286 times.
|
168556 | for (y = 0; y < size; y++) { |
123 |
2/2✓ Branch 0 taken 613504 times.
✓ Branch 1 taken 74992 times.
|
1376992 | for (x = 0; x < size; x++) { |
124 | 1227008 | *coeffs = *(uint16_t*)coeffs << -shift; | |
125 | 1227008 | coeffs++; | |
126 | } | ||
127 | } | ||
128 | } | ||
129 | 864282 | } | |
130 | |||
131 | #define SET(dst, x) (dst) = (x) | ||
132 | #define SCALE(dst, x) (dst) = av_clip_int16(((x) + add) >> shift) | ||
133 | |||
134 | #define TR_4x4_LUMA(dst, src, step, assign) \ | ||
135 | do { \ | ||
136 | int c0 = src[0 * step] + src[2 * step]; \ | ||
137 | int c1 = src[2 * step] + src[3 * step]; \ | ||
138 | int c2 = src[0 * step] - src[3 * step]; \ | ||
139 | int c3 = 74 * src[1 * step]; \ | ||
140 | \ | ||
141 | assign(dst[2 * step], 74 * (src[0 * step] - \ | ||
142 | src[2 * step] + \ | ||
143 | src[3 * step])); \ | ||
144 | assign(dst[0 * step], 29 * c0 + 55 * c1 + c3); \ | ||
145 | assign(dst[1 * step], 55 * c2 - 29 * c1 + c3); \ | ||
146 | assign(dst[3 * step], 55 * c0 + 29 * c2 - c3); \ | ||
147 | } while (0) | ||
148 | |||
149 | 7141354 | static void FUNC(transform_4x4_luma)(int16_t *coeffs) | |
150 | { | ||
151 | int i; | ||
152 | 7141354 | int shift = 7; | |
153 | 7141354 | int add = 1 << (shift - 1); | |
154 | 7141354 | int16_t *src = coeffs; | |
155 | |||
156 |
2/2✓ Branch 0 taken 14282708 times.
✓ Branch 1 taken 3570677 times.
|
35706770 | for (i = 0; i < 4; i++) { |
157 | 28565416 | TR_4x4_LUMA(src, src, 4, SCALE); | |
158 | 28565416 | src++; | |
159 | } | ||
160 | |||
161 | 7141354 | shift = 20 - BIT_DEPTH; | |
162 | 7141354 | add = 1 << (shift - 1); | |
163 |
2/2✓ Branch 0 taken 14282708 times.
✓ Branch 1 taken 3570677 times.
|
35706770 | for (i = 0; i < 4; i++) { |
164 | 28565416 | TR_4x4_LUMA(coeffs, coeffs, 1, SCALE); | |
165 | 28565416 | coeffs += 4; | |
166 | } | ||
167 | 7141354 | } | |
168 | |||
169 | #undef TR_4x4_LUMA | ||
170 | |||
171 | #define TR_4(dst, src, dstep, sstep, assign, end) \ | ||
172 | do { \ | ||
173 | const int e0 = 64 * src[0 * sstep] + 64 * src[2 * sstep]; \ | ||
174 | const int e1 = 64 * src[0 * sstep] - 64 * src[2 * sstep]; \ | ||
175 | const int o0 = 83 * src[1 * sstep] + 36 * src[3 * sstep]; \ | ||
176 | const int o1 = 36 * src[1 * sstep] - 83 * src[3 * sstep]; \ | ||
177 | \ | ||
178 | assign(dst[0 * dstep], e0 + o0); \ | ||
179 | assign(dst[1 * dstep], e1 + o1); \ | ||
180 | assign(dst[2 * dstep], e1 - o1); \ | ||
181 | assign(dst[3 * dstep], e0 - o0); \ | ||
182 | } while (0) | ||
183 | |||
184 | #define TR_8(dst, src, dstep, sstep, assign, end) \ | ||
185 | do { \ | ||
186 | int i, j; \ | ||
187 | int e_8[4]; \ | ||
188 | int o_8[4] = { 0 }; \ | ||
189 | for (i = 0; i < 4; i++) \ | ||
190 | for (j = 1; j < end; j += 2) \ | ||
191 | o_8[i] += transform[4 * j][i] * src[j * sstep]; \ | ||
192 | TR_4(e_8, src, 1, 2 * sstep, SET, 4); \ | ||
193 | \ | ||
194 | for (i = 0; i < 4; i++) { \ | ||
195 | assign(dst[i * dstep], e_8[i] + o_8[i]); \ | ||
196 | assign(dst[(7 - i) * dstep], e_8[i] - o_8[i]); \ | ||
197 | } \ | ||
198 | } while (0) | ||
199 | |||
200 | #define TR_16(dst, src, dstep, sstep, assign, end) \ | ||
201 | do { \ | ||
202 | int i, j; \ | ||
203 | int e_16[8]; \ | ||
204 | int o_16[8] = { 0 }; \ | ||
205 | for (i = 0; i < 8; i++) \ | ||
206 | for (j = 1; j < end; j += 2) \ | ||
207 | o_16[i] += transform[2 * j][i] * src[j * sstep]; \ | ||
208 | TR_8(e_16, src, 1, 2 * sstep, SET, 8); \ | ||
209 | \ | ||
210 | for (i = 0; i < 8; i++) { \ | ||
211 | assign(dst[i * dstep], e_16[i] + o_16[i]); \ | ||
212 | assign(dst[(15 - i) * dstep], e_16[i] - o_16[i]); \ | ||
213 | } \ | ||
214 | } while (0) | ||
215 | |||
216 | #define TR_32(dst, src, dstep, sstep, assign, end) \ | ||
217 | do { \ | ||
218 | int i, j; \ | ||
219 | int e_32[16]; \ | ||
220 | int o_32[16] = { 0 }; \ | ||
221 | for (i = 0; i < 16; i++) \ | ||
222 | for (j = 1; j < end; j += 2) \ | ||
223 | o_32[i] += transform[j][i] * src[j * sstep]; \ | ||
224 | TR_16(e_32, src, 1, 2 * sstep, SET, end / 2); \ | ||
225 | \ | ||
226 | for (i = 0; i < 16; i++) { \ | ||
227 | assign(dst[i * dstep], e_32[i] + o_32[i]); \ | ||
228 | assign(dst[(31 - i) * dstep], e_32[i] - o_32[i]); \ | ||
229 | } \ | ||
230 | } while (0) | ||
231 | |||
232 | #define IDCT_VAR4(H) \ | ||
233 | int limit2 = FFMIN(col_limit + 4, H) | ||
234 | #define IDCT_VAR8(H) \ | ||
235 | int limit = FFMIN(col_limit, H); \ | ||
236 | int limit2 = FFMIN(col_limit + 4, H) | ||
237 | #define IDCT_VAR16(H) IDCT_VAR8(H) | ||
238 | #define IDCT_VAR32(H) IDCT_VAR8(H) | ||
239 | |||
240 | #define IDCT(H) \ | ||
241 | static void FUNC(idct_ ## H ## x ## H )(int16_t *coeffs, \ | ||
242 | int col_limit) \ | ||
243 | { \ | ||
244 | int i; \ | ||
245 | int shift = 7; \ | ||
246 | int add = 1 << (shift - 1); \ | ||
247 | int16_t *src = coeffs; \ | ||
248 | IDCT_VAR ## H(H); \ | ||
249 | \ | ||
250 | for (i = 0; i < H; i++) { \ | ||
251 | TR_ ## H(src, src, H, H, SCALE, limit2); \ | ||
252 | if (limit2 < H && i%4 == 0 && !!i) \ | ||
253 | limit2 -= 4; \ | ||
254 | src++; \ | ||
255 | } \ | ||
256 | \ | ||
257 | shift = 20 - BIT_DEPTH; \ | ||
258 | add = 1 << (shift - 1); \ | ||
259 | for (i = 0; i < H; i++) { \ | ||
260 | TR_ ## H(coeffs, coeffs, 1, 1, SCALE, limit); \ | ||
261 | coeffs += H; \ | ||
262 | } \ | ||
263 | } | ||
264 | |||
265 | #define IDCT_DC(H) \ | ||
266 | static void FUNC(idct_ ## H ## x ## H ## _dc)(int16_t *coeffs) \ | ||
267 | { \ | ||
268 | int i, j; \ | ||
269 | int shift = 14 - BIT_DEPTH; \ | ||
270 | int add = 1 << (shift - 1); \ | ||
271 | int coeff = (((coeffs[0] + 1) >> 1) + add) >> shift; \ | ||
272 | \ | ||
273 | for (j = 0; j < H; j++) { \ | ||
274 | for (i = 0; i < H; i++) { \ | ||
275 | coeffs[i + j * H] = coeff; \ | ||
276 | } \ | ||
277 | } \ | ||
278 | } | ||
279 | |||
280 |
5/10✗ Branch 0 not taken.
✓ Branch 1 taken 10522240 times.
✗ Branch 2 not taken.
✗ Branch 3 not taken.
✗ Branch 4 not taken.
✗ Branch 5 not taken.
✓ Branch 6 taken 10522240 times.
✓ Branch 7 taken 2630560 times.
✓ Branch 8 taken 10522240 times.
✓ Branch 9 taken 2630560 times.
|
47350080 | IDCT( 4) |
281 |
17/22✓ Branch 0 taken 363635200 times.
✓ Branch 1 taken 90908800 times.
✓ Branch 2 taken 90908800 times.
✓ Branch 3 taken 22727200 times.
✓ Branch 4 taken 90908800 times.
✓ Branch 5 taken 22727200 times.
✗ Branch 6 not taken.
✓ Branch 7 taken 22727200 times.
✗ Branch 8 not taken.
✗ Branch 9 not taken.
✗ Branch 10 not taken.
✗ Branch 11 not taken.
✓ Branch 12 taken 22727200 times.
✓ Branch 13 taken 2840900 times.
✓ Branch 14 taken 257275712 times.
✓ Branch 15 taken 90908800 times.
✓ Branch 16 taken 90908800 times.
✓ Branch 17 taken 22727200 times.
✓ Branch 18 taken 90908800 times.
✓ Branch 19 taken 22727200 times.
✓ Branch 20 taken 22727200 times.
✓ Branch 21 taken 2840900 times.
|
2065682824 | IDCT( 8) |
282 |
34/34✓ Branch 0 taken 675830096 times.
✓ Branch 1 taken 177257856 times.
✓ Branch 2 taken 177257856 times.
✓ Branch 3 taken 22157232 times.
✓ Branch 4 taken 354515712 times.
✓ Branch 5 taken 88628928 times.
✓ Branch 6 taken 88628928 times.
✓ Branch 7 taken 22157232 times.
✓ Branch 8 taken 88628928 times.
✓ Branch 9 taken 22157232 times.
✓ Branch 10 taken 177257856 times.
✓ Branch 11 taken 22157232 times.
✓ Branch 12 taken 16376704 times.
✓ Branch 13 taken 5780528 times.
✓ Branch 14 taken 4094176 times.
✓ Branch 15 taken 12282528 times.
✓ Branch 16 taken 3070632 times.
✓ Branch 17 taken 1023544 times.
✓ Branch 18 taken 22157232 times.
✓ Branch 19 taken 1384827 times.
✓ Branch 20 taken 710084864 times.
✓ Branch 21 taken 177257856 times.
✓ Branch 22 taken 177257856 times.
✓ Branch 23 taken 22157232 times.
✓ Branch 24 taken 354515712 times.
✓ Branch 25 taken 88628928 times.
✓ Branch 26 taken 88628928 times.
✓ Branch 27 taken 22157232 times.
✓ Branch 28 taken 88628928 times.
✓ Branch 29 taken 22157232 times.
✓ Branch 30 taken 177257856 times.
✓ Branch 31 taken 22157232 times.
✓ Branch 32 taken 22157232 times.
✓ Branch 33 taken 1384827 times.
|
6408385622 | IDCT(16) |
283 |
46/46✓ Branch 0 taken 1116340416 times.
✓ Branch 1 taken 254356480 times.
✓ Branch 2 taken 254356480 times.
✓ Branch 3 taken 15897280 times.
✓ Branch 4 taken 271424888 times.
✓ Branch 5 taken 127178240 times.
✓ Branch 6 taken 127178240 times.
✓ Branch 7 taken 15897280 times.
✓ Branch 8 taken 254356480 times.
✓ Branch 9 taken 63589120 times.
✓ Branch 10 taken 63589120 times.
✓ Branch 11 taken 15897280 times.
✓ Branch 12 taken 63589120 times.
✓ Branch 13 taken 15897280 times.
✓ Branch 14 taken 127178240 times.
✓ Branch 15 taken 15897280 times.
✓ Branch 16 taken 254356480 times.
✓ Branch 17 taken 15897280 times.
✓ Branch 18 taken 13720416 times.
✓ Branch 19 taken 2176864 times.
✓ Branch 20 taken 3430104 times.
✓ Branch 21 taken 10290312 times.
✓ Branch 22 taken 3001341 times.
✓ Branch 23 taken 428763 times.
✓ Branch 24 taken 15897280 times.
✓ Branch 25 taken 496790 times.
✓ Branch 26 taken 1638708224 times.
✓ Branch 27 taken 254356480 times.
✓ Branch 28 taken 254356480 times.
✓ Branch 29 taken 15897280 times.
✓ Branch 30 taken 399364608 times.
✓ Branch 31 taken 127178240 times.
✓ Branch 32 taken 127178240 times.
✓ Branch 33 taken 15897280 times.
✓ Branch 34 taken 254356480 times.
✓ Branch 35 taken 63589120 times.
✓ Branch 36 taken 63589120 times.
✓ Branch 37 taken 15897280 times.
✓ Branch 38 taken 63589120 times.
✓ Branch 39 taken 15897280 times.
✓ Branch 40 taken 127178240 times.
✓ Branch 41 taken 15897280 times.
✓ Branch 42 taken 254356480 times.
✓ Branch 43 taken 15897280 times.
✓ Branch 44 taken 15897280 times.
✓ Branch 45 taken 496790 times.
|
11494675612 | IDCT(32) |
284 | |||
285 |
4/4✓ Branch 0 taken 17376480 times.
✓ Branch 1 taken 4344120 times.
✓ Branch 2 taken 4344120 times.
✓ Branch 3 taken 1086030 times.
|
45613260 | IDCT_DC( 4) |
286 |
4/4✓ Branch 0 taken 46935552 times.
✓ Branch 1 taken 5866944 times.
✓ Branch 2 taken 5866944 times.
✓ Branch 3 taken 733368 times.
|
107071728 | IDCT_DC( 8) |
287 |
4/4✓ Branch 0 taken 65159936 times.
✓ Branch 1 taken 4072496 times.
✓ Branch 2 taken 4072496 times.
✓ Branch 3 taken 254531 times.
|
138973926 | IDCT_DC(16) |
288 |
4/4✓ Branch 0 taken 79552512 times.
✓ Branch 1 taken 2486016 times.
✓ Branch 2 taken 2486016 times.
✓ Branch 3 taken 77688 times.
|
164232432 | IDCT_DC(32) |
289 | |||
290 | #undef TR_4 | ||
291 | #undef TR_8 | ||
292 | #undef TR_16 | ||
293 | #undef TR_32 | ||
294 | |||
295 | #undef SET | ||
296 | #undef SCALE | ||
297 | |||
298 | 222676 | static void FUNC(sao_band_filter)(uint8_t *_dst, const uint8_t *_src, | |
299 | ptrdiff_t stride_dst, ptrdiff_t stride_src, | ||
300 | const int16_t *sao_offset_val, int sao_left_class, | ||
301 | int width, int height) | ||
302 | { | ||
303 | 222676 | pixel *dst = (pixel *)_dst; | |
304 | 222676 | const pixel *src = (const pixel *)_src; | |
305 | 222676 | int offset_table[32] = { 0 }; | |
306 | int k, y, x; | ||
307 | 222676 | int shift = BIT_DEPTH - 5; | |
308 | |||
309 | 222676 | stride_dst /= sizeof(pixel); | |
310 | 222676 | stride_src /= sizeof(pixel); | |
311 | |||
312 |
2/2✓ Branch 0 taken 445352 times.
✓ Branch 1 taken 111338 times.
|
1113380 | for (k = 0; k < 4; k++) |
313 | 890704 | offset_table[(k + sao_left_class) & 31] = sao_offset_val[k + 1]; | |
314 |
2/2✓ Branch 0 taken 4185584 times.
✓ Branch 1 taken 111338 times.
|
8593844 | for (y = 0; y < height; y++) { |
315 |
2/2✓ Branch 0 taken 204816032 times.
✓ Branch 1 taken 4185584 times.
|
418003232 | for (x = 0; x < width; x++) |
316 | 409632064 | dst[x] = av_clip_pixel(src[x] + offset_table[(src[x] >> shift) & 31]); | |
317 | 8371168 | dst += stride_dst; | |
318 | 8371168 | src += stride_src; | |
319 | } | ||
320 | 222676 | } | |
321 | |||
322 | #define CMP(a, b) (((a) > (b)) - ((a) < (b))) | ||
323 | |||
324 | 744168 | static void FUNC(sao_edge_filter)(uint8_t *_dst, const uint8_t *_src, ptrdiff_t stride_dst, const int16_t *sao_offset_val, | |
325 | int eo, int width, int height) { | ||
326 | |||
327 | static const uint8_t edge_idx[] = { 1, 2, 0, 3, 4 }; | ||
328 | static const int8_t pos[4][2][2] = { | ||
329 | { { -1, 0 }, { 1, 0 } }, // horizontal | ||
330 | { { 0, -1 }, { 0, 1 } }, // vertical | ||
331 | { { -1, -1 }, { 1, 1 } }, // 45 degree | ||
332 | { { 1, -1 }, { -1, 1 } }, // 135 degree | ||
333 | }; | ||
334 | 744168 | pixel *dst = (pixel *)_dst; | |
335 | 744168 | const pixel *src = (const pixel *)_src; | |
336 | int a_stride, b_stride; | ||
337 | int x, y; | ||
338 | 744168 | ptrdiff_t stride_src = (2*MAX_PB_SIZE + AV_INPUT_BUFFER_PADDING_SIZE) / sizeof(pixel); | |
339 | 744168 | stride_dst /= sizeof(pixel); | |
340 | |||
341 | 744168 | a_stride = pos[eo][0][0] + pos[eo][0][1] * stride_src; | |
342 | 744168 | b_stride = pos[eo][1][0] + pos[eo][1][1] * stride_src; | |
343 |
2/2✓ Branch 0 taken 19188216 times.
✓ Branch 1 taken 372084 times.
|
39120600 | for (y = 0; y < height; y++) { |
344 |
2/2✓ Branch 0 taken 1057691808 times.
✓ Branch 1 taken 19188216 times.
|
2153760048 | for (x = 0; x < width; x++) { |
345 | 2115383616 | int diff0 = CMP(src[x], src[x + a_stride]); | |
346 | 2115383616 | int diff1 = CMP(src[x], src[x + b_stride]); | |
347 | 2115383616 | int offset_val = edge_idx[2 + diff0 + diff1]; | |
348 | 2115383616 | dst[x] = av_clip_pixel(src[x] + sao_offset_val[offset_val]); | |
349 | } | ||
350 | 38376432 | src += stride_src; | |
351 | 38376432 | dst += stride_dst; | |
352 | } | ||
353 | 744168 | } | |
354 | |||
355 | 721060 | static void FUNC(sao_edge_restore_0)(uint8_t *_dst, const uint8_t *_src, | |
356 | ptrdiff_t stride_dst, ptrdiff_t stride_src, const SAOParams *sao, | ||
357 | const int *borders, int _width, int _height, | ||
358 | int c_idx, const uint8_t *vert_edge, | ||
359 | const uint8_t *horiz_edge, const uint8_t *diag_edge) | ||
360 | { | ||
361 | int x, y; | ||
362 | 721060 | pixel *dst = (pixel *)_dst; | |
363 | 721060 | const pixel *src = (const pixel *)_src; | |
364 | 721060 | const int16_t *sao_offset_val = sao->offset_val[c_idx]; | |
365 | 721060 | int sao_eo_class = sao->eo_class[c_idx]; | |
366 | 721060 | int init_x = 0, width = _width, height = _height; | |
367 | |||
368 | 721060 | stride_dst /= sizeof(pixel); | |
369 | 721060 | stride_src /= sizeof(pixel); | |
370 | |||
371 |
2/2✓ Branch 0 taken 282668 times.
✓ Branch 1 taken 77862 times.
|
721060 | if (sao_eo_class != SAO_EO_VERT) { |
372 |
2/2✓ Branch 0 taken 13321 times.
✓ Branch 1 taken 269347 times.
|
565336 | if (borders[0]) { |
373 | 26642 | int offset_val = sao_offset_val[0]; | |
374 |
2/2✓ Branch 0 taken 657864 times.
✓ Branch 1 taken 13321 times.
|
1342370 | for (y = 0; y < height; y++) { |
375 | 1315728 | dst[y * stride_dst] = av_clip_pixel(src[y * stride_src] + offset_val); | |
376 | } | ||
377 | 26642 | init_x = 1; | |
378 | } | ||
379 |
2/2✓ Branch 0 taken 13553 times.
✓ Branch 1 taken 269115 times.
|
565336 | if (borders[2]) { |
380 | 27106 | int offset_val = sao_offset_val[0]; | |
381 | 27106 | int offset = width - 1; | |
382 |
2/2✓ Branch 0 taken 658904 times.
✓ Branch 1 taken 13553 times.
|
1344914 | for (x = 0; x < height; x++) { |
383 | 1317808 | dst[x * stride_dst + offset] = av_clip_pixel(src[x * stride_src + offset] + offset_val); | |
384 | } | ||
385 | 27106 | width--; | |
386 | } | ||
387 | } | ||
388 |
2/2✓ Branch 0 taken 296924 times.
✓ Branch 1 taken 63606 times.
|
721060 | if (sao_eo_class != SAO_EO_HORIZ) { |
389 |
2/2✓ Branch 0 taken 18461 times.
✓ Branch 1 taken 278463 times.
|
593848 | if (borders[1]) { |
390 | 36922 | int offset_val = sao_offset_val[0]; | |
391 |
2/2✓ Branch 0 taken 907223 times.
✓ Branch 1 taken 18461 times.
|
1851368 | for (x = init_x; x < width; x++) |
392 | 1814446 | dst[x] = av_clip_pixel(src[x] + offset_val); | |
393 | } | ||
394 |
2/2✓ Branch 0 taken 23393 times.
✓ Branch 1 taken 273531 times.
|
593848 | if (borders[3]) { |
395 | 46786 | int offset_val = sao_offset_val[0]; | |
396 | 46786 | ptrdiff_t y_stride_dst = stride_dst * (height - 1); | |
397 | 46786 | ptrdiff_t y_stride_src = stride_src * (height - 1); | |
398 |
2/2✓ Branch 0 taken 1115478 times.
✓ Branch 1 taken 23393 times.
|
2277742 | for (x = init_x; x < width; x++) |
399 | 2230956 | dst[x + y_stride_dst] = av_clip_pixel(src[x + y_stride_src] + offset_val); | |
400 | 46786 | height--; | |
401 | } | ||
402 | } | ||
403 | 721060 | } | |
404 | |||
405 | 38834 | static void FUNC(sao_edge_restore_1)(uint8_t *_dst, const uint8_t *_src, | |
406 | ptrdiff_t stride_dst, ptrdiff_t stride_src, const SAOParams *sao, | ||
407 | const int *borders, int _width, int _height, | ||
408 | int c_idx, const uint8_t *vert_edge, | ||
409 | const uint8_t *horiz_edge, const uint8_t *diag_edge) | ||
410 | { | ||
411 | int x, y; | ||
412 | 38834 | pixel *dst = (pixel *)_dst; | |
413 | 38834 | const pixel *src = (const pixel *)_src; | |
414 | 38834 | const int16_t *sao_offset_val = sao->offset_val[c_idx]; | |
415 | 38834 | int sao_eo_class = sao->eo_class[c_idx]; | |
416 | 38834 | int init_x = 0, init_y = 0, width = _width, height = _height; | |
417 | |||
418 | 38834 | stride_dst /= sizeof(pixel); | |
419 | 38834 | stride_src /= sizeof(pixel); | |
420 | |||
421 |
2/2✓ Branch 0 taken 13613 times.
✓ Branch 1 taken 5804 times.
|
38834 | if (sao_eo_class != SAO_EO_VERT) { |
422 |
2/2✓ Branch 0 taken 1252 times.
✓ Branch 1 taken 12361 times.
|
27226 | if (borders[0]) { |
423 | 2504 | int offset_val = sao_offset_val[0]; | |
424 |
2/2✓ Branch 0 taken 61472 times.
✓ Branch 1 taken 1252 times.
|
125448 | for (y = 0; y < height; y++) { |
425 | 122944 | dst[y * stride_dst] = av_clip_pixel(src[y * stride_src] + offset_val); | |
426 | } | ||
427 | 2504 | init_x = 1; | |
428 | } | ||
429 |
2/2✓ Branch 0 taken 795 times.
✓ Branch 1 taken 12818 times.
|
27226 | if (borders[2]) { |
430 | 1590 | int offset_val = sao_offset_val[0]; | |
431 | 1590 | int offset = width - 1; | |
432 |
2/2✓ Branch 0 taken 37304 times.
✓ Branch 1 taken 795 times.
|
76198 | for (x = 0; x < height; x++) { |
433 | 74608 | dst[x * stride_dst + offset] = av_clip_pixel(src[x * stride_src + offset] + offset_val); | |
434 | } | ||
435 | 1590 | width--; | |
436 | } | ||
437 | } | ||
438 |
2/2✓ Branch 0 taken 13993 times.
✓ Branch 1 taken 5424 times.
|
38834 | if (sao_eo_class != SAO_EO_HORIZ) { |
439 |
2/2✓ Branch 0 taken 1963 times.
✓ Branch 1 taken 12030 times.
|
27986 | if (borders[1]) { |
440 | 3926 | int offset_val = sao_offset_val[0]; | |
441 |
2/2✓ Branch 0 taken 94362 times.
✓ Branch 1 taken 1963 times.
|
192650 | for (x = init_x; x < width; x++) |
442 | 188724 | dst[x] = av_clip_pixel(src[x] + offset_val); | |
443 | 3926 | init_y = 1; | |
444 | } | ||
445 |
2/2✓ Branch 0 taken 1584 times.
✓ Branch 1 taken 12409 times.
|
27986 | if (borders[3]) { |
446 | 3168 | int offset_val = sao_offset_val[0]; | |
447 | 3168 | ptrdiff_t y_stride_dst = stride_dst * (height - 1); | |
448 | 3168 | ptrdiff_t y_stride_src = stride_src * (height - 1); | |
449 |
2/2✓ Branch 0 taken 73429 times.
✓ Branch 1 taken 1584 times.
|
150026 | for (x = init_x; x < width; x++) |
450 | 146858 | dst[x + y_stride_dst] = av_clip_pixel(src[x + y_stride_src] + offset_val); | |
451 | 3168 | height--; | |
452 | } | ||
453 | } | ||
454 | |||
455 | { | ||
456 |
8/8✓ Branch 0 taken 19253 times.
✓ Branch 1 taken 164 times.
✓ Branch 2 taken 3639 times.
✓ Branch 3 taken 15614 times.
✓ Branch 4 taken 3342 times.
✓ Branch 5 taken 297 times.
✓ Branch 6 taken 3092 times.
✓ Branch 7 taken 250 times.
|
38834 | int save_upper_left = !diag_edge[0] && sao_eo_class == SAO_EO_135D && !borders[0] && !borders[1]; |
457 |
8/8✓ Branch 0 taken 19205 times.
✓ Branch 1 taken 212 times.
✓ Branch 2 taken 4388 times.
✓ Branch 3 taken 14817 times.
✓ Branch 4 taken 3661 times.
✓ Branch 5 taken 727 times.
✓ Branch 6 taken 3437 times.
✓ Branch 7 taken 224 times.
|
38834 | int save_upper_right = !diag_edge[1] && sao_eo_class == SAO_EO_45D && !borders[1] && !borders[2]; |
458 |
8/8✓ Branch 0 taken 19103 times.
✓ Branch 1 taken 314 times.
✓ Branch 2 taken 3587 times.
✓ Branch 3 taken 15516 times.
✓ Branch 4 taken 3388 times.
✓ Branch 5 taken 199 times.
✓ Branch 6 taken 2943 times.
✓ Branch 7 taken 445 times.
|
38834 | int save_lower_right = !diag_edge[2] && sao_eo_class == SAO_EO_135D && !borders[2] && !borders[3]; |
459 |
8/8✓ Branch 0 taken 19155 times.
✓ Branch 1 taken 262 times.
✓ Branch 2 taken 4410 times.
✓ Branch 3 taken 14745 times.
✓ Branch 4 taken 3852 times.
✓ Branch 5 taken 558 times.
✓ Branch 6 taken 3392 times.
✓ Branch 7 taken 460 times.
|
38834 | int save_lower_left = !diag_edge[3] && sao_eo_class == SAO_EO_45D && !borders[0] && !borders[3]; |
460 | |||
461 | // Restore pixels that can't be modified | ||
462 |
4/4✓ Branch 0 taken 46 times.
✓ Branch 1 taken 19371 times.
✓ Branch 2 taken 38 times.
✓ Branch 3 taken 8 times.
|
38834 | if(vert_edge[0] && sao_eo_class != SAO_EO_VERT) { |
463 |
2/2✓ Branch 0 taken 1216 times.
✓ Branch 1 taken 38 times.
|
2508 | for(y = init_y+save_upper_left; y< height-save_lower_left; y++) |
464 | 2432 | dst[y*stride_dst] = src[y*stride_src]; | |
465 | } | ||
466 |
3/4✓ Branch 0 taken 92 times.
✓ Branch 1 taken 19325 times.
✓ Branch 2 taken 92 times.
✗ Branch 3 not taken.
|
38834 | if(vert_edge[1] && sao_eo_class != SAO_EO_VERT) { |
467 |
2/2✓ Branch 0 taken 2812 times.
✓ Branch 1 taken 92 times.
|
5808 | for(y = init_y+save_upper_right; y< height-save_lower_right; y++) |
468 | 5624 | dst[y*stride_dst+width-1] = src[y*stride_src+width-1]; | |
469 | } | ||
470 | |||
471 |
4/4✓ Branch 0 taken 130 times.
✓ Branch 1 taken 19287 times.
✓ Branch 2 taken 116 times.
✓ Branch 3 taken 14 times.
|
38834 | if(horiz_edge[0] && sao_eo_class != SAO_EO_HORIZ) { |
472 |
2/2✓ Branch 0 taken 3700 times.
✓ Branch 1 taken 116 times.
|
7632 | for(x = init_x+save_upper_left; x < width-save_upper_right; x++) |
473 | 7400 | dst[x] = src[x]; | |
474 | } | ||
475 |
4/4✓ Branch 0 taken 234 times.
✓ Branch 1 taken 19183 times.
✓ Branch 2 taken 180 times.
✓ Branch 3 taken 54 times.
|
38834 | if(horiz_edge[1] && sao_eo_class != SAO_EO_HORIZ) { |
476 |
2/2✓ Branch 0 taken 5742 times.
✓ Branch 1 taken 180 times.
|
11844 | for(x = init_x+save_lower_left; x < width-save_lower_right; x++) |
477 | 11484 | dst[(height-1)*stride_dst+x] = src[(height-1)*stride_src+x]; | |
478 | } | ||
479 |
4/4✓ Branch 0 taken 164 times.
✓ Branch 1 taken 19253 times.
✓ Branch 2 taken 56 times.
✓ Branch 3 taken 108 times.
|
38834 | if(diag_edge[0] && sao_eo_class == SAO_EO_135D) |
480 | 112 | dst[0] = src[0]; | |
481 |
4/4✓ Branch 0 taken 212 times.
✓ Branch 1 taken 19205 times.
✓ Branch 2 taken 106 times.
✓ Branch 3 taken 106 times.
|
38834 | if(diag_edge[1] && sao_eo_class == SAO_EO_45D) |
482 | 212 | dst[width-1] = src[width-1]; | |
483 |
4/4✓ Branch 0 taken 314 times.
✓ Branch 1 taken 19103 times.
✓ Branch 2 taken 108 times.
✓ Branch 3 taken 206 times.
|
38834 | if(diag_edge[2] && sao_eo_class == SAO_EO_135D) |
484 | 216 | dst[stride_dst*(height-1)+width-1] = src[stride_src*(height-1)+width-1]; | |
485 |
4/4✓ Branch 0 taken 262 times.
✓ Branch 1 taken 19155 times.
✓ Branch 2 taken 84 times.
✓ Branch 3 taken 178 times.
|
38834 | if(diag_edge[3] && sao_eo_class == SAO_EO_45D) |
486 | 168 | dst[stride_dst*(height-1)] = src[stride_src*(height-1)]; | |
487 | |||
488 | } | ||
489 | 38834 | } | |
490 | |||
491 | #undef CMP | ||
492 | |||
493 | //////////////////////////////////////////////////////////////////////////////// | ||
494 | // | ||
495 | //////////////////////////////////////////////////////////////////////////////// | ||
496 | 4751490 | static void FUNC(put_hevc_pel_pixels)(int16_t *dst, | |
497 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
498 | int height, intptr_t mx, intptr_t my, int width) | ||
499 | { | ||
500 | int x, y; | ||
501 | 4751490 | const pixel *src = (const pixel *)_src; | |
502 | 4751490 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
503 | |||
504 |
2/2✓ Branch 0 taken 35314934 times.
✓ Branch 1 taken 2375745 times.
|
75381358 | for (y = 0; y < height; y++) { |
505 |
2/2✓ Branch 0 taken 898412916 times.
✓ Branch 1 taken 35314934 times.
|
1867455700 | for (x = 0; x < width; x++) |
506 | 1796825832 | dst[x] = src[x] << (14 - BIT_DEPTH); | |
507 | 70629868 | src += srcstride; | |
508 | 70629868 | dst += MAX_PB_SIZE; | |
509 | } | ||
510 | 4751490 | } | |
511 | |||
512 | 4783994 | static void FUNC(put_hevc_pel_uni_pixels)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
513 | int height, intptr_t mx, intptr_t my, int width) | ||
514 | { | ||
515 | int y; | ||
516 | 4783994 | const pixel *src = (const pixel *)_src; | |
517 | 4783994 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
518 | 4783994 | pixel *dst = (pixel *)_dst; | |
519 | 4783994 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
520 | |||
521 |
2/2✓ Branch 0 taken 25285382 times.
✓ Branch 1 taken 2391997 times.
|
55354758 | for (y = 0; y < height; y++) { |
522 | 50570764 | memcpy(dst, src, width * sizeof(pixel)); | |
523 | 50570764 | src += srcstride; | |
524 | 50570764 | dst += dststride; | |
525 | } | ||
526 | 4783994 | } | |
527 | |||
528 | 5198258 | static void FUNC(put_hevc_pel_bi_pixels)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
529 | const int16_t *src2, | ||
530 | int height, intptr_t mx, intptr_t my, int width) | ||
531 | { | ||
532 | int x, y; | ||
533 | 5198258 | const pixel *src = (const pixel *)_src; | |
534 | 5198258 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
535 | 5198258 | pixel *dst = (pixel *)_dst; | |
536 | 5198258 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
537 | |||
538 | 5198258 | int shift = 14 + 1 - BIT_DEPTH; | |
539 | #if BIT_DEPTH < 14 | ||
540 | 5198258 | int offset = 1 << (shift - 1); | |
541 | #else | ||
542 | int offset = 0; | ||
543 | #endif | ||
544 | |||
545 |
2/2✓ Branch 0 taken 38401358 times.
✓ Branch 1 taken 2599129 times.
|
82000974 | for (y = 0; y < height; y++) { |
546 |
2/2✓ Branch 0 taken 967620084 times.
✓ Branch 1 taken 38401358 times.
|
2012042884 | for (x = 0; x < width; x++) |
547 | 1935240168 | dst[x] = av_clip_pixel(((src[x] << (14 - BIT_DEPTH)) + src2[x] + offset) >> shift); | |
548 | 76802716 | src += srcstride; | |
549 | 76802716 | dst += dststride; | |
550 | 76802716 | src2 += MAX_PB_SIZE; | |
551 | } | ||
552 | 5198258 | } | |
553 | |||
554 | 145958 | static void FUNC(put_hevc_pel_uni_w_pixels)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
555 | int height, int denom, int wx, int ox, intptr_t mx, intptr_t my, int width) | ||
556 | { | ||
557 | int x, y; | ||
558 | 145958 | const pixel *src = (const pixel *)_src; | |
559 | 145958 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
560 | 145958 | pixel *dst = (pixel *)_dst; | |
561 | 145958 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
562 | 145958 | int shift = denom + 14 - BIT_DEPTH; | |
563 | #if BIT_DEPTH < 14 | ||
564 | 145958 | int offset = 1 << (shift - 1); | |
565 | #else | ||
566 | int offset = 0; | ||
567 | #endif | ||
568 | |||
569 | 145958 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
570 |
2/2✓ Branch 0 taken 1640336 times.
✓ Branch 1 taken 72979 times.
|
3426630 | for (y = 0; y < height; y++) { |
571 |
2/2✓ Branch 0 taken 55907336 times.
✓ Branch 1 taken 1640336 times.
|
115095344 | for (x = 0; x < width; x++) |
572 | 111814672 | dst[x] = av_clip_pixel((((src[x] << (14 - BIT_DEPTH)) * wx + offset) >> shift) + ox); | |
573 | 3280672 | src += srcstride; | |
574 | 3280672 | dst += dststride; | |
575 | } | ||
576 | 145958 | } | |
577 | |||
578 | 64154 | static void FUNC(put_hevc_pel_bi_w_pixels)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
579 | const int16_t *src2, | ||
580 | int height, int denom, int wx0, int wx1, | ||
581 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
582 | { | ||
583 | int x, y; | ||
584 | 64154 | const pixel *src = (const pixel *)_src; | |
585 | 64154 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
586 | 64154 | pixel *dst = (pixel *)_dst; | |
587 | 64154 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
588 | |||
589 | 64154 | int shift = 14 + 1 - BIT_DEPTH; | |
590 | 64154 | int log2Wd = denom + shift - 1; | |
591 | |||
592 | 64154 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
593 | 64154 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
594 |
2/2✓ Branch 0 taken 671000 times.
✓ Branch 1 taken 32077 times.
|
1406154 | for (y = 0; y < height; y++) { |
595 |
2/2✓ Branch 0 taken 21890728 times.
✓ Branch 1 taken 671000 times.
|
45123456 | for (x = 0; x < width; x++) { |
596 | 43781456 | dst[x] = av_clip_pixel(( (src[x] << (14 - BIT_DEPTH)) * wx1 + src2[x] * wx0 + (ox0 + ox1 + 1) * (1 << log2Wd)) >> (log2Wd + 1)); | |
597 | } | ||
598 | 1342000 | src += srcstride; | |
599 | 1342000 | dst += dststride; | |
600 | 1342000 | src2 += MAX_PB_SIZE; | |
601 | } | ||
602 | 64154 | } | |
603 | |||
604 | //////////////////////////////////////////////////////////////////////////////// | ||
605 | // | ||
606 | //////////////////////////////////////////////////////////////////////////////// | ||
607 | #define QPEL_FILTER(src, stride) \ | ||
608 | (filter[0] * src[x - 3 * stride] + \ | ||
609 | filter[1] * src[x - 2 * stride] + \ | ||
610 | filter[2] * src[x - stride] + \ | ||
611 | filter[3] * src[x ] + \ | ||
612 | filter[4] * src[x + stride] + \ | ||
613 | filter[5] * src[x + 2 * stride] + \ | ||
614 | filter[6] * src[x + 3 * stride] + \ | ||
615 | filter[7] * src[x + 4 * stride]) | ||
616 | |||
617 | 1367194 | static void FUNC(put_hevc_qpel_h)(int16_t *dst, | |
618 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
619 | int height, intptr_t mx, intptr_t my, int width) | ||
620 | { | ||
621 | int x, y; | ||
622 | 1367194 | const pixel *src = (const pixel*)_src; | |
623 | 1367194 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
624 | 1367194 | const int8_t *filter = ff_hevc_qpel_filters[mx - 1]; | |
625 |
2/2✓ Branch 0 taken 14792996 times.
✓ Branch 1 taken 683597 times.
|
30953186 | for (y = 0; y < height; y++) { |
626 |
2/2✓ Branch 0 taken 486842632 times.
✓ Branch 1 taken 14792996 times.
|
1003271256 | for (x = 0; x < width; x++) |
627 | 973685264 | dst[x] = QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
628 | 29585992 | src += srcstride; | |
629 | 29585992 | dst += MAX_PB_SIZE; | |
630 | } | ||
631 | 1367194 | } | |
632 | |||
633 | 1070912 | static void FUNC(put_hevc_qpel_v)(int16_t *dst, | |
634 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
635 | int height, intptr_t mx, intptr_t my, int width) | ||
636 | { | ||
637 | int x, y; | ||
638 | 1070912 | const pixel *src = (const pixel*)_src; | |
639 | 1070912 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
640 | 1070912 | const int8_t *filter = ff_hevc_qpel_filters[my - 1]; | |
641 |
2/2✓ Branch 0 taken 11550632 times.
✓ Branch 1 taken 535456 times.
|
24172176 | for (y = 0; y < height; y++) { |
642 |
2/2✓ Branch 0 taken 379349896 times.
✓ Branch 1 taken 11550632 times.
|
781801056 | for (x = 0; x < width; x++) |
643 | 758699792 | dst[x] = QPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8); | |
644 | 23101264 | src += srcstride; | |
645 | 23101264 | dst += MAX_PB_SIZE; | |
646 | } | ||
647 | 1070912 | } | |
648 | |||
649 | 3342990 | static void FUNC(put_hevc_qpel_hv)(int16_t *dst, | |
650 | const uint8_t *_src, | ||
651 | ptrdiff_t _srcstride, | ||
652 | int height, intptr_t mx, | ||
653 | intptr_t my, int width) | ||
654 | { | ||
655 | int x, y; | ||
656 | const int8_t *filter; | ||
657 | 3342990 | const pixel *src = (const pixel*)_src; | |
658 | 3342990 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
659 | int16_t tmp_array[(MAX_PB_SIZE + QPEL_EXTRA) * MAX_PB_SIZE]; | ||
660 | 3342990 | int16_t *tmp = tmp_array; | |
661 | |||
662 | 3342990 | src -= QPEL_EXTRA_BEFORE * srcstride; | |
663 | 3342990 | filter = ff_hevc_qpel_filters[mx - 1]; | |
664 |
2/2✓ Branch 0 taken 46886537 times.
✓ Branch 1 taken 1671495 times.
|
97116064 | for (y = 0; y < height + QPEL_EXTRA; y++) { |
665 |
2/2✓ Branch 0 taken 1349183996 times.
✓ Branch 1 taken 46886537 times.
|
2792141066 | for (x = 0; x < width; x++) |
666 | 2698367992 | tmp[x] = QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
667 | 93773074 | src += srcstride; | |
668 | 93773074 | tmp += MAX_PB_SIZE; | |
669 | } | ||
670 | |||
671 | 3342990 | tmp = tmp_array + QPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
672 | 3342990 | filter = ff_hevc_qpel_filters[my - 1]; | |
673 |
2/2✓ Branch 0 taken 35186072 times.
✓ Branch 1 taken 1671495 times.
|
73715134 | for (y = 0; y < height; y++) { |
674 |
2/2✓ Branch 0 taken 1105803656 times.
✓ Branch 1 taken 35186072 times.
|
2281979456 | for (x = 0; x < width; x++) |
675 | 2211607312 | dst[x] = QPEL_FILTER(tmp, MAX_PB_SIZE) >> 6; | |
676 | 70372144 | tmp += MAX_PB_SIZE; | |
677 | 70372144 | dst += MAX_PB_SIZE; | |
678 | } | ||
679 | 3342990 | } | |
680 | |||
681 | 1570408 | static void FUNC(put_hevc_qpel_uni_h)(uint8_t *_dst, ptrdiff_t _dststride, | |
682 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
683 | int height, intptr_t mx, intptr_t my, int width) | ||
684 | { | ||
685 | int x, y; | ||
686 | 1570408 | const pixel *src = (const pixel*)_src; | |
687 | 1570408 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
688 | 1570408 | pixel *dst = (pixel *)_dst; | |
689 | 1570408 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
690 | 1570408 | const int8_t *filter = ff_hevc_qpel_filters[mx - 1]; | |
691 | 1570408 | int shift = 14 - BIT_DEPTH; | |
692 | |||
693 | #if BIT_DEPTH < 14 | ||
694 | 1570408 | int offset = 1 << (shift - 1); | |
695 | #else | ||
696 | int offset = 0; | ||
697 | #endif | ||
698 | |||
699 |
2/2✓ Branch 0 taken 11363824 times.
✓ Branch 1 taken 785204 times.
|
24298056 | for (y = 0; y < height; y++) { |
700 |
2/2✓ Branch 0 taken 247342760 times.
✓ Branch 1 taken 11363824 times.
|
517413168 | for (x = 0; x < width; x++) |
701 | 494685520 | dst[x] = av_clip_pixel(((QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) + offset) >> shift); | |
702 | 22727648 | src += srcstride; | |
703 | 22727648 | dst += dststride; | |
704 | } | ||
705 | 1570408 | } | |
706 | |||
707 | 1287348 | static void FUNC(put_hevc_qpel_bi_h)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
708 | const int16_t *src2, | ||
709 | int height, intptr_t mx, intptr_t my, int width) | ||
710 | { | ||
711 | int x, y; | ||
712 | 1287348 | const pixel *src = (const pixel*)_src; | |
713 | 1287348 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
714 | 1287348 | pixel *dst = (pixel *)_dst; | |
715 | 1287348 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
716 | |||
717 | 1287348 | const int8_t *filter = ff_hevc_qpel_filters[mx - 1]; | |
718 | |||
719 | 1287348 | int shift = 14 + 1 - BIT_DEPTH; | |
720 | #if BIT_DEPTH < 14 | ||
721 | 1287348 | int offset = 1 << (shift - 1); | |
722 | #else | ||
723 | int offset = 0; | ||
724 | #endif | ||
725 | |||
726 |
2/2✓ Branch 0 taken 13936548 times.
✓ Branch 1 taken 643674 times.
|
29160444 | for (y = 0; y < height; y++) { |
727 |
2/2✓ Branch 0 taken 457729096 times.
✓ Branch 1 taken 13936548 times.
|
943331288 | for (x = 0; x < width; x++) |
728 | 915458192 | dst[x] = av_clip_pixel(((QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) + src2[x] + offset) >> shift); | |
729 | 27873096 | src += srcstride; | |
730 | 27873096 | dst += dststride; | |
731 | 27873096 | src2 += MAX_PB_SIZE; | |
732 | } | ||
733 | 1287348 | } | |
734 | |||
735 | 1371102 | static void FUNC(put_hevc_qpel_uni_v)(uint8_t *_dst, ptrdiff_t _dststride, | |
736 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
737 | int height, intptr_t mx, intptr_t my, int width) | ||
738 | { | ||
739 | int x, y; | ||
740 | 1371102 | const pixel *src = (const pixel*)_src; | |
741 | 1371102 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
742 | 1371102 | pixel *dst = (pixel *)_dst; | |
743 | 1371102 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
744 | 1371102 | const int8_t *filter = ff_hevc_qpel_filters[my - 1]; | |
745 | 1371102 | int shift = 14 - BIT_DEPTH; | |
746 | |||
747 | #if BIT_DEPTH < 14 | ||
748 | 1371102 | int offset = 1 << (shift - 1); | |
749 | #else | ||
750 | int offset = 0; | ||
751 | #endif | ||
752 | |||
753 |
2/2✓ Branch 0 taken 10197500 times.
✓ Branch 1 taken 685551 times.
|
21766102 | for (y = 0; y < height; y++) { |
754 |
2/2✓ Branch 0 taken 237113384 times.
✓ Branch 1 taken 10197500 times.
|
494621768 | for (x = 0; x < width; x++) |
755 | 474226768 | dst[x] = av_clip_pixel(((QPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) + offset) >> shift); | |
756 | 20395000 | src += srcstride; | |
757 | 20395000 | dst += dststride; | |
758 | } | ||
759 | 1371102 | } | |
760 | |||
761 | |||
762 | 1024872 | static void FUNC(put_hevc_qpel_bi_v)(uint8_t *_dst, ptrdiff_t _dststride, | |
763 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
764 | int height, intptr_t mx, intptr_t my, int width) | ||
765 | { | ||
766 | int x, y; | ||
767 | 1024872 | const pixel *src = (const pixel*)_src; | |
768 | 1024872 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
769 | 1024872 | pixel *dst = (pixel *)_dst; | |
770 | 1024872 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
771 | |||
772 | 1024872 | const int8_t *filter = ff_hevc_qpel_filters[my - 1]; | |
773 | |||
774 | 1024872 | int shift = 14 + 1 - BIT_DEPTH; | |
775 | #if BIT_DEPTH < 14 | ||
776 | 1024872 | int offset = 1 << (shift - 1); | |
777 | #else | ||
778 | int offset = 0; | ||
779 | #endif | ||
780 | |||
781 |
2/2✓ Branch 0 taken 10864452 times.
✓ Branch 1 taken 512436 times.
|
22753776 | for (y = 0; y < height; y++) { |
782 |
2/2✓ Branch 0 taken 348786952 times.
✓ Branch 1 taken 10864452 times.
|
719302808 | for (x = 0; x < width; x++) |
783 | 697573904 | dst[x] = av_clip_pixel(((QPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) + src2[x] + offset) >> shift); | |
784 | 21728904 | src += srcstride; | |
785 | 21728904 | dst += dststride; | |
786 | 21728904 | src2 += MAX_PB_SIZE; | |
787 | } | ||
788 | 1024872 | } | |
789 | |||
790 | 4327134 | static void FUNC(put_hevc_qpel_uni_hv)(uint8_t *_dst, ptrdiff_t _dststride, | |
791 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
792 | int height, intptr_t mx, intptr_t my, int width) | ||
793 | { | ||
794 | int x, y; | ||
795 | const int8_t *filter; | ||
796 | 4327134 | const pixel *src = (const pixel*)_src; | |
797 | 4327134 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
798 | 4327134 | pixel *dst = (pixel *)_dst; | |
799 | 4327134 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
800 | int16_t tmp_array[(MAX_PB_SIZE + QPEL_EXTRA) * MAX_PB_SIZE]; | ||
801 | 4327134 | int16_t *tmp = tmp_array; | |
802 | 4327134 | int shift = 14 - BIT_DEPTH; | |
803 | |||
804 | #if BIT_DEPTH < 14 | ||
805 | 4327134 | int offset = 1 << (shift - 1); | |
806 | #else | ||
807 | int offset = 0; | ||
808 | #endif | ||
809 | |||
810 | 4327134 | src -= QPEL_EXTRA_BEFORE * srcstride; | |
811 | 4327134 | filter = ff_hevc_qpel_filters[mx - 1]; | |
812 |
2/2✓ Branch 0 taken 46357257 times.
✓ Branch 1 taken 2163567 times.
|
97041648 | for (y = 0; y < height + QPEL_EXTRA; y++) { |
813 |
2/2✓ Branch 0 taken 871399616 times.
✓ Branch 1 taken 46357257 times.
|
1835513746 | for (x = 0; x < width; x++) |
814 | 1742799232 | tmp[x] = QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
815 | 92714514 | src += srcstride; | |
816 | 92714514 | tmp += MAX_PB_SIZE; | |
817 | } | ||
818 | |||
819 | 4327134 | tmp = tmp_array + QPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
820 | 4327134 | filter = ff_hevc_qpel_filters[my - 1]; | |
821 | |||
822 |
2/2✓ Branch 0 taken 31212288 times.
✓ Branch 1 taken 2163567 times.
|
66751710 | for (y = 0; y < height; y++) { |
823 |
2/2✓ Branch 0 taken 658027688 times.
✓ Branch 1 taken 31212288 times.
|
1378479952 | for (x = 0; x < width; x++) |
824 | 1316055376 | dst[x] = av_clip_pixel(((QPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) + offset) >> shift); | |
825 | 62424576 | tmp += MAX_PB_SIZE; | |
826 | 62424576 | dst += dststride; | |
827 | } | ||
828 | 4327134 | } | |
829 | |||
830 | 3195682 | static void FUNC(put_hevc_qpel_bi_hv)(uint8_t *_dst, ptrdiff_t _dststride, | |
831 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
832 | int height, intptr_t mx, intptr_t my, int width) | ||
833 | { | ||
834 | int x, y; | ||
835 | const int8_t *filter; | ||
836 | 3195682 | const pixel *src = (const pixel*)_src; | |
837 | 3195682 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
838 | 3195682 | pixel *dst = (pixel *)_dst; | |
839 | 3195682 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
840 | int16_t tmp_array[(MAX_PB_SIZE + QPEL_EXTRA) * MAX_PB_SIZE]; | ||
841 | 3195682 | int16_t *tmp = tmp_array; | |
842 | 3195682 | int shift = 14 + 1 - BIT_DEPTH; | |
843 | #if BIT_DEPTH < 14 | ||
844 | 3195682 | int offset = 1 << (shift - 1); | |
845 | #else | ||
846 | int offset = 0; | ||
847 | #endif | ||
848 | |||
849 | 3195682 | src -= QPEL_EXTRA_BEFORE * srcstride; | |
850 | 3195682 | filter = ff_hevc_qpel_filters[mx - 1]; | |
851 |
2/2✓ Branch 0 taken 44950087 times.
✓ Branch 1 taken 1597841 times.
|
93095856 | for (y = 0; y < height + QPEL_EXTRA; y++) { |
852 |
2/2✓ Branch 0 taken 1304439384 times.
✓ Branch 1 taken 44950087 times.
|
2698778942 | for (x = 0; x < width; x++) |
853 | 2608878768 | tmp[x] = QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
854 | 89900174 | src += srcstride; | |
855 | 89900174 | tmp += MAX_PB_SIZE; | |
856 | } | ||
857 | |||
858 | 3195682 | tmp = tmp_array + QPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
859 | 3195682 | filter = ff_hevc_qpel_filters[my - 1]; | |
860 | |||
861 |
2/2✓ Branch 0 taken 33765200 times.
✓ Branch 1 taken 1597841 times.
|
70726082 | for (y = 0; y < height; y++) { |
862 |
2/2✓ Branch 0 taken 1070725512 times.
✓ Branch 1 taken 33765200 times.
|
2208981424 | for (x = 0; x < width; x++) |
863 | 2141451024 | dst[x] = av_clip_pixel(((QPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) + src2[x] + offset) >> shift); | |
864 | 67530400 | tmp += MAX_PB_SIZE; | |
865 | 67530400 | dst += dststride; | |
866 | 67530400 | src2 += MAX_PB_SIZE; | |
867 | } | ||
868 | 3195682 | } | |
869 | |||
870 | 36230 | static void FUNC(put_hevc_qpel_uni_w_h)(uint8_t *_dst, ptrdiff_t _dststride, | |
871 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
872 | int height, int denom, int wx, int ox, | ||
873 | intptr_t mx, intptr_t my, int width) | ||
874 | { | ||
875 | int x, y; | ||
876 | 36230 | const pixel *src = (const pixel*)_src; | |
877 | 36230 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
878 | 36230 | pixel *dst = (pixel *)_dst; | |
879 | 36230 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
880 | 36230 | const int8_t *filter = ff_hevc_qpel_filters[mx - 1]; | |
881 | 36230 | int shift = denom + 14 - BIT_DEPTH; | |
882 | #if BIT_DEPTH < 14 | ||
883 | 36230 | int offset = 1 << (shift - 1); | |
884 | #else | ||
885 | int offset = 0; | ||
886 | #endif | ||
887 | |||
888 | 36230 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
889 |
2/2✓ Branch 0 taken 381616 times.
✓ Branch 1 taken 18115 times.
|
799462 | for (y = 0; y < height; y++) { |
890 |
2/2✓ Branch 0 taken 12448016 times.
✓ Branch 1 taken 381616 times.
|
25659264 | for (x = 0; x < width; x++) |
891 | 24896032 | dst[x] = av_clip_pixel((((QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox); | |
892 | 763232 | src += srcstride; | |
893 | 763232 | dst += dststride; | |
894 | } | ||
895 | 36230 | } | |
896 | |||
897 | 22810 | static void FUNC(put_hevc_qpel_bi_w_h)(uint8_t *_dst, ptrdiff_t _dststride, | |
898 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
899 | int height, int denom, int wx0, int wx1, | ||
900 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
901 | { | ||
902 | int x, y; | ||
903 | 22810 | const pixel *src = (const pixel*)_src; | |
904 | 22810 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
905 | 22810 | pixel *dst = (pixel *)_dst; | |
906 | 22810 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
907 | |||
908 | 22810 | const int8_t *filter = ff_hevc_qpel_filters[mx - 1]; | |
909 | |||
910 | 22810 | int shift = 14 + 1 - BIT_DEPTH; | |
911 | 22810 | int log2Wd = denom + shift - 1; | |
912 | |||
913 | 22810 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
914 | 22810 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
915 |
2/2✓ Branch 0 taken 278600 times.
✓ Branch 1 taken 11405 times.
|
580010 | for (y = 0; y < height; y++) { |
916 |
2/2✓ Branch 0 taken 9869200 times.
✓ Branch 1 taken 278600 times.
|
20295600 | for (x = 0; x < width; x++) |
917 | 19738400 | dst[x] = av_clip_pixel(((QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) * wx1 + src2[x] * wx0 + | |
918 | ((ox0 + ox1 + 1) * (1 << log2Wd))) >> (log2Wd + 1)); | ||
919 | 557200 | src += srcstride; | |
920 | 557200 | dst += dststride; | |
921 | 557200 | src2 += MAX_PB_SIZE; | |
922 | } | ||
923 | 22810 | } | |
924 | |||
925 | 32950 | static void FUNC(put_hevc_qpel_uni_w_v)(uint8_t *_dst, ptrdiff_t _dststride, | |
926 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
927 | int height, int denom, int wx, int ox, | ||
928 | intptr_t mx, intptr_t my, int width) | ||
929 | { | ||
930 | int x, y; | ||
931 | 32950 | const pixel *src = (const pixel*)_src; | |
932 | 32950 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
933 | 32950 | pixel *dst = (pixel *)_dst; | |
934 | 32950 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
935 | 32950 | const int8_t *filter = ff_hevc_qpel_filters[my - 1]; | |
936 | 32950 | int shift = denom + 14 - BIT_DEPTH; | |
937 | #if BIT_DEPTH < 14 | ||
938 | 32950 | int offset = 1 << (shift - 1); | |
939 | #else | ||
940 | int offset = 0; | ||
941 | #endif | ||
942 | |||
943 | 32950 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
944 |
2/2✓ Branch 0 taken 318884 times.
✓ Branch 1 taken 16475 times.
|
670718 | for (y = 0; y < height; y++) { |
945 |
2/2✓ Branch 0 taken 9467248 times.
✓ Branch 1 taken 318884 times.
|
19572264 | for (x = 0; x < width; x++) |
946 | 18934496 | dst[x] = av_clip_pixel((((QPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox); | |
947 | 637768 | src += srcstride; | |
948 | 637768 | dst += dststride; | |
949 | } | ||
950 | 32950 | } | |
951 | |||
952 | 20448 | static void FUNC(put_hevc_qpel_bi_w_v)(uint8_t *_dst, ptrdiff_t _dststride, | |
953 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
954 | int height, int denom, int wx0, int wx1, | ||
955 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
956 | { | ||
957 | int x, y; | ||
958 | 20448 | const pixel *src = (const pixel*)_src; | |
959 | 20448 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
960 | 20448 | pixel *dst = (pixel *)_dst; | |
961 | 20448 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
962 | |||
963 | 20448 | const int8_t *filter = ff_hevc_qpel_filters[my - 1]; | |
964 | |||
965 | 20448 | int shift = 14 + 1 - BIT_DEPTH; | |
966 | 20448 | int log2Wd = denom + shift - 1; | |
967 | |||
968 | 20448 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
969 | 20448 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
970 |
2/2✓ Branch 0 taken 236324 times.
✓ Branch 1 taken 10224 times.
|
493096 | for (y = 0; y < height; y++) { |
971 |
2/2✓ Branch 0 taken 7942032 times.
✓ Branch 1 taken 236324 times.
|
16356712 | for (x = 0; x < width; x++) |
972 | 15884064 | dst[x] = av_clip_pixel(((QPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) * wx1 + src2[x] * wx0 + | |
973 | ((ox0 + ox1 + 1) * (1 << log2Wd))) >> (log2Wd + 1)); | ||
974 | 472648 | src += srcstride; | |
975 | 472648 | dst += dststride; | |
976 | 472648 | src2 += MAX_PB_SIZE; | |
977 | } | ||
978 | 20448 | } | |
979 | |||
980 | 102878 | static void FUNC(put_hevc_qpel_uni_w_hv)(uint8_t *_dst, ptrdiff_t _dststride, | |
981 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
982 | int height, int denom, int wx, int ox, | ||
983 | intptr_t mx, intptr_t my, int width) | ||
984 | { | ||
985 | int x, y; | ||
986 | const int8_t *filter; | ||
987 | 102878 | const pixel *src = (const pixel*)_src; | |
988 | 102878 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
989 | 102878 | pixel *dst = (pixel *)_dst; | |
990 | 102878 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
991 | int16_t tmp_array[(MAX_PB_SIZE + QPEL_EXTRA) * MAX_PB_SIZE]; | ||
992 | 102878 | int16_t *tmp = tmp_array; | |
993 | 102878 | int shift = denom + 14 - BIT_DEPTH; | |
994 | #if BIT_DEPTH < 14 | ||
995 | 102878 | int offset = 1 << (shift - 1); | |
996 | #else | ||
997 | int offset = 0; | ||
998 | #endif | ||
999 | |||
1000 | 102878 | src -= QPEL_EXTRA_BEFORE * srcstride; | |
1001 | 102878 | filter = ff_hevc_qpel_filters[mx - 1]; | |
1002 |
2/2✓ Branch 0 taken 1372533 times.
✓ Branch 1 taken 51439 times.
|
2847944 | for (y = 0; y < height + QPEL_EXTRA; y++) { |
1003 |
2/2✓ Branch 0 taken 37487844 times.
✓ Branch 1 taken 1372533 times.
|
77720754 | for (x = 0; x < width; x++) |
1004 | 74975688 | tmp[x] = QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1005 | 2745066 | src += srcstride; | |
1006 | 2745066 | tmp += MAX_PB_SIZE; | |
1007 | } | ||
1008 | |||
1009 | 102878 | tmp = tmp_array + QPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1010 | 102878 | filter = ff_hevc_qpel_filters[my - 1]; | |
1011 | |||
1012 | 102878 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
1013 |
2/2✓ Branch 0 taken 1012460 times.
✓ Branch 1 taken 51439 times.
|
2127798 | for (y = 0; y < height; y++) { |
1014 |
2/2✓ Branch 0 taken 30718032 times.
✓ Branch 1 taken 1012460 times.
|
63460984 | for (x = 0; x < width; x++) |
1015 | 61436064 | dst[x] = av_clip_pixel((((QPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) * wx + offset) >> shift) + ox); | |
1016 | 2024920 | tmp += MAX_PB_SIZE; | |
1017 | 2024920 | dst += dststride; | |
1018 | } | ||
1019 | 102878 | } | |
1020 | |||
1021 | 71042 | static void FUNC(put_hevc_qpel_bi_w_hv)(uint8_t *_dst, ptrdiff_t _dststride, | |
1022 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1023 | int height, int denom, int wx0, int wx1, | ||
1024 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
1025 | { | ||
1026 | int x, y; | ||
1027 | const int8_t *filter; | ||
1028 | 71042 | const pixel *src = (const pixel*)_src; | |
1029 | 71042 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1030 | 71042 | pixel *dst = (pixel *)_dst; | |
1031 | 71042 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1032 | int16_t tmp_array[(MAX_PB_SIZE + QPEL_EXTRA) * MAX_PB_SIZE]; | ||
1033 | 71042 | int16_t *tmp = tmp_array; | |
1034 | 71042 | int shift = 14 + 1 - BIT_DEPTH; | |
1035 | 71042 | int log2Wd = denom + shift - 1; | |
1036 | |||
1037 | 71042 | src -= QPEL_EXTRA_BEFORE * srcstride; | |
1038 | 71042 | filter = ff_hevc_qpel_filters[mx - 1]; | |
1039 |
2/2✓ Branch 0 taken 1131207 times.
✓ Branch 1 taken 35521 times.
|
2333456 | for (y = 0; y < height + QPEL_EXTRA; y++) { |
1040 |
2/2✓ Branch 0 taken 36956072 times.
✓ Branch 1 taken 1131207 times.
|
76174558 | for (x = 0; x < width; x++) |
1041 | 73912144 | tmp[x] = QPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1042 | 2262414 | src += srcstride; | |
1043 | 2262414 | tmp += MAX_PB_SIZE; | |
1044 | } | ||
1045 | |||
1046 | 71042 | tmp = tmp_array + QPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1047 | 71042 | filter = ff_hevc_qpel_filters[my - 1]; | |
1048 | |||
1049 | 71042 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
1050 | 71042 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
1051 |
2/2✓ Branch 0 taken 882560 times.
✓ Branch 1 taken 35521 times.
|
1836162 | for (y = 0; y < height; y++) { |
1052 |
2/2✓ Branch 0 taken 30984848 times.
✓ Branch 1 taken 882560 times.
|
63734816 | for (x = 0; x < width; x++) |
1053 | 61969696 | dst[x] = av_clip_pixel(((QPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) * wx1 + src2[x] * wx0 + | |
1054 | ((ox0 + ox1 + 1) * (1 << log2Wd))) >> (log2Wd + 1)); | ||
1055 | 1765120 | tmp += MAX_PB_SIZE; | |
1056 | 1765120 | dst += dststride; | |
1057 | 1765120 | src2 += MAX_PB_SIZE; | |
1058 | } | ||
1059 | 71042 | } | |
1060 | |||
1061 | //////////////////////////////////////////////////////////////////////////////// | ||
1062 | // | ||
1063 | //////////////////////////////////////////////////////////////////////////////// | ||
1064 | #define EPEL_FILTER(src, stride) \ | ||
1065 | (filter[0] * src[x - stride] + \ | ||
1066 | filter[1] * src[x] + \ | ||
1067 | filter[2] * src[x + stride] + \ | ||
1068 | filter[3] * src[x + 2 * stride]) | ||
1069 | |||
1070 | 2148602 | static void FUNC(put_hevc_epel_h)(int16_t *dst, | |
1071 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
1072 | int height, intptr_t mx, intptr_t my, int width) | ||
1073 | { | ||
1074 | int x, y; | ||
1075 | 2148602 | const pixel *src = (const pixel *)_src; | |
1076 | 2148602 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1077 | 2148602 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1078 |
2/2✓ Branch 0 taken 12458290 times.
✓ Branch 1 taken 1074301 times.
|
27065182 | for (y = 0; y < height; y++) { |
1079 |
2/2✓ Branch 0 taken 201156308 times.
✓ Branch 1 taken 12458290 times.
|
427229196 | for (x = 0; x < width; x++) |
1080 | 402312616 | dst[x] = EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1081 | 24916580 | src += srcstride; | |
1082 | 24916580 | dst += MAX_PB_SIZE; | |
1083 | } | ||
1084 | 2148602 | } | |
1085 | |||
1086 | 1491966 | static void FUNC(put_hevc_epel_v)(int16_t *dst, | |
1087 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
1088 | int height, intptr_t mx, intptr_t my, int width) | ||
1089 | { | ||
1090 | int x, y; | ||
1091 | 1491966 | const pixel *src = (const pixel *)_src; | |
1092 | 1491966 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1093 | 1491966 | const int8_t *filter = ff_hevc_epel_filters[my - 1]; | |
1094 | |||
1095 |
2/2✓ Branch 0 taken 9097742 times.
✓ Branch 1 taken 745983 times.
|
19687450 | for (y = 0; y < height; y++) { |
1096 |
2/2✓ Branch 0 taken 154103796 times.
✓ Branch 1 taken 9097742 times.
|
326403076 | for (x = 0; x < width; x++) |
1097 | 308207592 | dst[x] = EPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8); | |
1098 | 18195484 | src += srcstride; | |
1099 | 18195484 | dst += MAX_PB_SIZE; | |
1100 | } | ||
1101 | 1491966 | } | |
1102 | |||
1103 | 8511430 | static void FUNC(put_hevc_epel_hv)(int16_t *dst, | |
1104 | const uint8_t *_src, ptrdiff_t _srcstride, | ||
1105 | int height, intptr_t mx, intptr_t my, int width) | ||
1106 | { | ||
1107 | int x, y; | ||
1108 | 8511430 | const pixel *src = (const pixel *)_src; | |
1109 | 8511430 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1110 | 8511430 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1111 | int16_t tmp_array[(MAX_PB_SIZE + EPEL_EXTRA) * MAX_PB_SIZE]; | ||
1112 | 8511430 | int16_t *tmp = tmp_array; | |
1113 | |||
1114 | 8511430 | src -= EPEL_EXTRA_BEFORE * srcstride; | |
1115 | |||
1116 |
2/2✓ Branch 0 taken 60681831 times.
✓ Branch 1 taken 4255715 times.
|
129875092 | for (y = 0; y < height + EPEL_EXTRA; y++) { |
1117 |
2/2✓ Branch 0 taken 886091878 times.
✓ Branch 1 taken 60681831 times.
|
1893547418 | for (x = 0; x < width; x++) |
1118 | 1772183756 | tmp[x] = EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1119 | 121363662 | src += srcstride; | |
1120 | 121363662 | tmp += MAX_PB_SIZE; | |
1121 | } | ||
1122 | |||
1123 | 8511430 | tmp = tmp_array + EPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1124 | 8511430 | filter = ff_hevc_epel_filters[my - 1]; | |
1125 | |||
1126 |
2/2✓ Branch 0 taken 47914686 times.
✓ Branch 1 taken 4255715 times.
|
104340802 | for (y = 0; y < height; y++) { |
1127 |
2/2✓ Branch 0 taken 751566868 times.
✓ Branch 1 taken 47914686 times.
|
1598963108 | for (x = 0; x < width; x++) |
1128 | 1503133736 | dst[x] = EPEL_FILTER(tmp, MAX_PB_SIZE) >> 6; | |
1129 | 95829372 | tmp += MAX_PB_SIZE; | |
1130 | 95829372 | dst += MAX_PB_SIZE; | |
1131 | } | ||
1132 | 8511430 | } | |
1133 | |||
1134 | 2564074 | static void FUNC(put_hevc_epel_uni_h)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
1135 | int height, intptr_t mx, intptr_t my, int width) | ||
1136 | { | ||
1137 | int x, y; | ||
1138 | 2564074 | const pixel *src = (const pixel *)_src; | |
1139 | 2564074 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1140 | 2564074 | pixel *dst = (pixel *)_dst; | |
1141 | 2564074 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1142 | 2564074 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1143 | 2564074 | int shift = 14 - BIT_DEPTH; | |
1144 | #if BIT_DEPTH < 14 | ||
1145 | 2564074 | int offset = 1 << (shift - 1); | |
1146 | #else | ||
1147 | int offset = 0; | ||
1148 | #endif | ||
1149 | |||
1150 |
2/2✓ Branch 0 taken 9225530 times.
✓ Branch 1 taken 1282037 times.
|
21015134 | for (y = 0; y < height; y++) { |
1151 |
2/2✓ Branch 0 taken 97805140 times.
✓ Branch 1 taken 9225530 times.
|
214061340 | for (x = 0; x < width; x++) |
1152 | 195610280 | dst[x] = av_clip_pixel(((EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) + offset) >> shift); | |
1153 | 18451060 | src += srcstride; | |
1154 | 18451060 | dst += dststride; | |
1155 | } | ||
1156 | 2564074 | } | |
1157 | |||
1158 | 2020670 | static void FUNC(put_hevc_epel_bi_h)(uint8_t *_dst, ptrdiff_t _dststride, | |
1159 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1160 | int height, intptr_t mx, intptr_t my, int width) | ||
1161 | { | ||
1162 | int x, y; | ||
1163 | 2020670 | const pixel *src = (const pixel *)_src; | |
1164 | 2020670 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1165 | 2020670 | pixel *dst = (pixel *)_dst; | |
1166 | 2020670 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1167 | 2020670 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1168 | 2020670 | int shift = 14 + 1 - BIT_DEPTH; | |
1169 | #if BIT_DEPTH < 14 | ||
1170 | 2020670 | int offset = 1 << (shift - 1); | |
1171 | #else | ||
1172 | int offset = 0; | ||
1173 | #endif | ||
1174 | |||
1175 |
2/2✓ Branch 0 taken 11523918 times.
✓ Branch 1 taken 1010335 times.
|
25068506 | for (y = 0; y < height; y++) { |
1176 |
2/2✓ Branch 0 taken 186679348 times.
✓ Branch 1 taken 11523918 times.
|
396406532 | for (x = 0; x < width; x++) { |
1177 | 373358696 | dst[x] = av_clip_pixel(((EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) + src2[x] + offset) >> shift); | |
1178 | } | ||
1179 | 23047836 | dst += dststride; | |
1180 | 23047836 | src += srcstride; | |
1181 | 23047836 | src2 += MAX_PB_SIZE; | |
1182 | } | ||
1183 | 2020670 | } | |
1184 | |||
1185 | 2092498 | static void FUNC(put_hevc_epel_uni_v)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
1186 | int height, intptr_t mx, intptr_t my, int width) | ||
1187 | { | ||
1188 | int x, y; | ||
1189 | 2092498 | const pixel *src = (const pixel *)_src; | |
1190 | 2092498 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1191 | 2092498 | pixel *dst = (pixel *)_dst; | |
1192 | 2092498 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1193 | 2092498 | const int8_t *filter = ff_hevc_epel_filters[my - 1]; | |
1194 | 2092498 | int shift = 14 - BIT_DEPTH; | |
1195 | #if BIT_DEPTH < 14 | ||
1196 | 2092498 | int offset = 1 << (shift - 1); | |
1197 | #else | ||
1198 | int offset = 0; | ||
1199 | #endif | ||
1200 | |||
1201 |
2/2✓ Branch 0 taken 7802770 times.
✓ Branch 1 taken 1046249 times.
|
17698038 | for (y = 0; y < height; y++) { |
1202 |
2/2✓ Branch 0 taken 89203556 times.
✓ Branch 1 taken 7802770 times.
|
194012652 | for (x = 0; x < width; x++) |
1203 | 178407112 | dst[x] = av_clip_pixel(((EPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) + offset) >> shift); | |
1204 | 15605540 | src += srcstride; | |
1205 | 15605540 | dst += dststride; | |
1206 | } | ||
1207 | 2092498 | } | |
1208 | |||
1209 | 1448070 | static void FUNC(put_hevc_epel_bi_v)(uint8_t *_dst, ptrdiff_t _dststride, | |
1210 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1211 | int height, intptr_t mx, intptr_t my, int width) | ||
1212 | { | ||
1213 | int x, y; | ||
1214 | 1448070 | const pixel *src = (const pixel *)_src; | |
1215 | 1448070 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1216 | 1448070 | const int8_t *filter = ff_hevc_epel_filters[my - 1]; | |
1217 | 1448070 | pixel *dst = (pixel *)_dst; | |
1218 | 1448070 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1219 | 1448070 | int shift = 14 + 1 - BIT_DEPTH; | |
1220 | #if BIT_DEPTH < 14 | ||
1221 | 1448070 | int offset = 1 << (shift - 1); | |
1222 | #else | ||
1223 | int offset = 0; | ||
1224 | #endif | ||
1225 | |||
1226 |
2/2✓ Branch 0 taken 8683602 times.
✓ Branch 1 taken 724035 times.
|
18815274 | for (y = 0; y < height; y++) { |
1227 |
2/2✓ Branch 0 taken 141364564 times.
✓ Branch 1 taken 8683602 times.
|
300096332 | for (x = 0; x < width; x++) |
1228 | 282729128 | dst[x] = av_clip_pixel(((EPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) + src2[x] + offset) >> shift); | |
1229 | 17367204 | dst += dststride; | |
1230 | 17367204 | src += srcstride; | |
1231 | 17367204 | src2 += MAX_PB_SIZE; | |
1232 | } | ||
1233 | 1448070 | } | |
1234 | |||
1235 | 11438578 | static void FUNC(put_hevc_epel_uni_hv)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
1236 | int height, intptr_t mx, intptr_t my, int width) | ||
1237 | { | ||
1238 | int x, y; | ||
1239 | 11438578 | const pixel *src = (const pixel *)_src; | |
1240 | 11438578 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1241 | 11438578 | pixel *dst = (pixel *)_dst; | |
1242 | 11438578 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1243 | 11438578 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1244 | int16_t tmp_array[(MAX_PB_SIZE + EPEL_EXTRA) * MAX_PB_SIZE]; | ||
1245 | 11438578 | int16_t *tmp = tmp_array; | |
1246 | 11438578 | int shift = 14 - BIT_DEPTH; | |
1247 | #if BIT_DEPTH < 14 | ||
1248 | 11438578 | int offset = 1 << (shift - 1); | |
1249 | #else | ||
1250 | int offset = 0; | ||
1251 | #endif | ||
1252 | |||
1253 | 11438578 | src -= EPEL_EXTRA_BEFORE * srcstride; | |
1254 | |||
1255 |
2/2✓ Branch 0 taken 58795917 times.
✓ Branch 1 taken 5719289 times.
|
129030412 | for (y = 0; y < height + EPEL_EXTRA; y++) { |
1256 |
2/2✓ Branch 0 taken 563719898 times.
✓ Branch 1 taken 58795917 times.
|
1245031630 | for (x = 0; x < width; x++) |
1257 | 1127439796 | tmp[x] = EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1258 | 117591834 | src += srcstride; | |
1259 | 117591834 | tmp += MAX_PB_SIZE; | |
1260 | } | ||
1261 | |||
1262 | 11438578 | tmp = tmp_array + EPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1263 | 11438578 | filter = ff_hevc_epel_filters[my - 1]; | |
1264 | |||
1265 |
2/2✓ Branch 0 taken 41638050 times.
✓ Branch 1 taken 5719289 times.
|
94714678 | for (y = 0; y < height; y++) { |
1266 |
2/2✓ Branch 0 taken 442624772 times.
✓ Branch 1 taken 41638050 times.
|
968525644 | for (x = 0; x < width; x++) |
1267 | 885249544 | dst[x] = av_clip_pixel(((EPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) + offset) >> shift); | |
1268 | 83276100 | tmp += MAX_PB_SIZE; | |
1269 | 83276100 | dst += dststride; | |
1270 | } | ||
1271 | 11438578 | } | |
1272 | |||
1273 | 8135842 | static void FUNC(put_hevc_epel_bi_hv)(uint8_t *_dst, ptrdiff_t _dststride, | |
1274 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1275 | int height, intptr_t mx, intptr_t my, int width) | ||
1276 | { | ||
1277 | int x, y; | ||
1278 | 8135842 | const pixel *src = (const pixel *)_src; | |
1279 | 8135842 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1280 | 8135842 | pixel *dst = (pixel *)_dst; | |
1281 | 8135842 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1282 | 8135842 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1283 | int16_t tmp_array[(MAX_PB_SIZE + EPEL_EXTRA) * MAX_PB_SIZE]; | ||
1284 | 8135842 | int16_t *tmp = tmp_array; | |
1285 | 8135842 | int shift = 14 + 1 - BIT_DEPTH; | |
1286 | #if BIT_DEPTH < 14 | ||
1287 | 8135842 | int offset = 1 << (shift - 1); | |
1288 | #else | ||
1289 | int offset = 0; | ||
1290 | #endif | ||
1291 | |||
1292 | 8135842 | src -= EPEL_EXTRA_BEFORE * srcstride; | |
1293 | |||
1294 |
2/2✓ Branch 0 taken 58184829 times.
✓ Branch 1 taken 4067921 times.
|
124505500 | for (y = 0; y < height + EPEL_EXTRA; y++) { |
1295 |
2/2✓ Branch 0 taken 848106018 times.
✓ Branch 1 taken 58184829 times.
|
1812581694 | for (x = 0; x < width; x++) |
1296 | 1696212036 | tmp[x] = EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1297 | 116369658 | src += srcstride; | |
1298 | 116369658 | tmp += MAX_PB_SIZE; | |
1299 | } | ||
1300 | |||
1301 | 8135842 | tmp = tmp_array + EPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1302 | 8135842 | filter = ff_hevc_epel_filters[my - 1]; | |
1303 | |||
1304 |
2/2✓ Branch 0 taken 45981066 times.
✓ Branch 1 taken 4067921 times.
|
100097974 | for (y = 0; y < height; y++) { |
1305 |
2/2✓ Branch 0 taken 719577684 times.
✓ Branch 1 taken 45981066 times.
|
1531117500 | for (x = 0; x < width; x++) |
1306 | 1439155368 | dst[x] = av_clip_pixel(((EPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) + src2[x] + offset) >> shift); | |
1307 | 91962132 | tmp += MAX_PB_SIZE; | |
1308 | 91962132 | dst += dststride; | |
1309 | 91962132 | src2 += MAX_PB_SIZE; | |
1310 | } | ||
1311 | 8135842 | } | |
1312 | |||
1313 | 53580 | static void FUNC(put_hevc_epel_uni_w_h)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
1314 | int height, int denom, int wx, int ox, intptr_t mx, intptr_t my, int width) | ||
1315 | { | ||
1316 | int x, y; | ||
1317 | 53580 | const pixel *src = (const pixel *)_src; | |
1318 | 53580 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1319 | 53580 | pixel *dst = (pixel *)_dst; | |
1320 | 53580 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1321 | 53580 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1322 | 53580 | int shift = denom + 14 - BIT_DEPTH; | |
1323 | #if BIT_DEPTH < 14 | ||
1324 | 53580 | int offset = 1 << (shift - 1); | |
1325 | #else | ||
1326 | int offset = 0; | ||
1327 | #endif | ||
1328 | |||
1329 | 53580 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
1330 |
2/2✓ Branch 0 taken 323716 times.
✓ Branch 1 taken 26790 times.
|
701012 | for (y = 0; y < height; y++) { |
1331 |
2/2✓ Branch 0 taken 6563880 times.
✓ Branch 1 taken 323716 times.
|
13775192 | for (x = 0; x < width; x++) { |
1332 | 13127760 | dst[x] = av_clip_pixel((((EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox); | |
1333 | } | ||
1334 | 647432 | dst += dststride; | |
1335 | 647432 | src += srcstride; | |
1336 | } | ||
1337 | 53580 | } | |
1338 | |||
1339 | 33080 | static void FUNC(put_hevc_epel_bi_w_h)(uint8_t *_dst, ptrdiff_t _dststride, | |
1340 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1341 | int height, int denom, int wx0, int wx1, | ||
1342 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
1343 | { | ||
1344 | int x, y; | ||
1345 | 33080 | const pixel *src = (const pixel *)_src; | |
1346 | 33080 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1347 | 33080 | pixel *dst = (pixel *)_dst; | |
1348 | 33080 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1349 | 33080 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1350 | 33080 | int shift = 14 + 1 - BIT_DEPTH; | |
1351 | 33080 | int log2Wd = denom + shift - 1; | |
1352 | |||
1353 | 33080 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
1354 | 33080 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
1355 |
2/2✓ Branch 0 taken 233792 times.
✓ Branch 1 taken 16540 times.
|
500664 | for (y = 0; y < height; y++) { |
1356 |
2/2✓ Branch 0 taken 5306152 times.
✓ Branch 1 taken 233792 times.
|
11079888 | for (x = 0; x < width; x++) |
1357 | 10612304 | dst[x] = av_clip_pixel(((EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8)) * wx1 + src2[x] * wx0 + | |
1358 | ((ox0 + ox1 + 1) * (1 << log2Wd))) >> (log2Wd + 1)); | ||
1359 | 467584 | src += srcstride; | |
1360 | 467584 | dst += dststride; | |
1361 | 467584 | src2 += MAX_PB_SIZE; | |
1362 | } | ||
1363 | 33080 | } | |
1364 | |||
1365 | 42000 | static void FUNC(put_hevc_epel_uni_w_v)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
1366 | int height, int denom, int wx, int ox, intptr_t mx, intptr_t my, int width) | ||
1367 | { | ||
1368 | int x, y; | ||
1369 | 42000 | const pixel *src = (const pixel *)_src; | |
1370 | 42000 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1371 | 42000 | pixel *dst = (pixel *)_dst; | |
1372 | 42000 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1373 | 42000 | const int8_t *filter = ff_hevc_epel_filters[my - 1]; | |
1374 | 42000 | int shift = denom + 14 - BIT_DEPTH; | |
1375 | #if BIT_DEPTH < 14 | ||
1376 | 42000 | int offset = 1 << (shift - 1); | |
1377 | #else | ||
1378 | int offset = 0; | ||
1379 | #endif | ||
1380 | |||
1381 | 42000 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
1382 |
2/2✓ Branch 0 taken 236812 times.
✓ Branch 1 taken 21000 times.
|
515624 | for (y = 0; y < height; y++) { |
1383 |
2/2✓ Branch 0 taken 4764408 times.
✓ Branch 1 taken 236812 times.
|
10002440 | for (x = 0; x < width; x++) { |
1384 | 9528816 | dst[x] = av_clip_pixel((((EPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) * wx + offset) >> shift) + ox); | |
1385 | } | ||
1386 | 473624 | dst += dststride; | |
1387 | 473624 | src += srcstride; | |
1388 | } | ||
1389 | 42000 | } | |
1390 | |||
1391 | 24552 | static void FUNC(put_hevc_epel_bi_w_v)(uint8_t *_dst, ptrdiff_t _dststride, | |
1392 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1393 | int height, int denom, int wx0, int wx1, | ||
1394 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
1395 | { | ||
1396 | int x, y; | ||
1397 | 24552 | const pixel *src = (const pixel *)_src; | |
1398 | 24552 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1399 | 24552 | const int8_t *filter = ff_hevc_epel_filters[my - 1]; | |
1400 | 24552 | pixel *dst = (pixel *)_dst; | |
1401 | 24552 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1402 | 24552 | int shift = 14 + 1 - BIT_DEPTH; | |
1403 | 24552 | int log2Wd = denom + shift - 1; | |
1404 | |||
1405 | 24552 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
1406 | 24552 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
1407 |
2/2✓ Branch 0 taken 167956 times.
✓ Branch 1 taken 12276 times.
|
360464 | for (y = 0; y < height; y++) { |
1408 |
2/2✓ Branch 0 taken 3922696 times.
✓ Branch 1 taken 167956 times.
|
8181304 | for (x = 0; x < width; x++) |
1409 | 7845392 | dst[x] = av_clip_pixel(((EPEL_FILTER(src, srcstride) >> (BIT_DEPTH - 8)) * wx1 + src2[x] * wx0 + | |
1410 | ((ox0 + ox1 + 1) * (1 << log2Wd))) >> (log2Wd + 1)); | ||
1411 | 335912 | src += srcstride; | |
1412 | 335912 | dst += dststride; | |
1413 | 335912 | src2 += MAX_PB_SIZE; | |
1414 | } | ||
1415 | 24552 | } | |
1416 | |||
1417 | 251200 | static void FUNC(put_hevc_epel_uni_w_hv)(uint8_t *_dst, ptrdiff_t _dststride, const uint8_t *_src, ptrdiff_t _srcstride, | |
1418 | int height, int denom, int wx, int ox, intptr_t mx, intptr_t my, int width) | ||
1419 | { | ||
1420 | int x, y; | ||
1421 | 251200 | const pixel *src = (const pixel *)_src; | |
1422 | 251200 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1423 | 251200 | pixel *dst = (pixel *)_dst; | |
1424 | 251200 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1425 | 251200 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1426 | int16_t tmp_array[(MAX_PB_SIZE + EPEL_EXTRA) * MAX_PB_SIZE]; | ||
1427 | 251200 | int16_t *tmp = tmp_array; | |
1428 | 251200 | int shift = denom + 14 - BIT_DEPTH; | |
1429 | #if BIT_DEPTH < 14 | ||
1430 | 251200 | int offset = 1 << (shift - 1); | |
1431 | #else | ||
1432 | int offset = 0; | ||
1433 | #endif | ||
1434 | |||
1435 | 251200 | src -= EPEL_EXTRA_BEFORE * srcstride; | |
1436 | |||
1437 |
2/2✓ Branch 0 taken 1597200 times.
✓ Branch 1 taken 125600 times.
|
3445600 | for (y = 0; y < height + EPEL_EXTRA; y++) { |
1438 |
2/2✓ Branch 0 taken 22400688 times.
✓ Branch 1 taken 1597200 times.
|
47995776 | for (x = 0; x < width; x++) |
1439 | 44801376 | tmp[x] = EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1440 | 3194400 | src += srcstride; | |
1441 | 3194400 | tmp += MAX_PB_SIZE; | |
1442 | } | ||
1443 | |||
1444 | 251200 | tmp = tmp_array + EPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1445 | 251200 | filter = ff_hevc_epel_filters[my - 1]; | |
1446 | |||
1447 | 251200 | ox = ox * (1 << (BIT_DEPTH - 8)); | |
1448 |
2/2✓ Branch 0 taken 1220400 times.
✓ Branch 1 taken 125600 times.
|
2692000 | for (y = 0; y < height; y++) { |
1449 |
2/2✓ Branch 0 taken 18910776 times.
✓ Branch 1 taken 1220400 times.
|
40262352 | for (x = 0; x < width; x++) |
1450 | 37821552 | dst[x] = av_clip_pixel((((EPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) * wx + offset) >> shift) + ox); | |
1451 | 2440800 | tmp += MAX_PB_SIZE; | |
1452 | 2440800 | dst += dststride; | |
1453 | } | ||
1454 | 251200 | } | |
1455 | |||
1456 | 166916 | static void FUNC(put_hevc_epel_bi_w_hv)(uint8_t *_dst, ptrdiff_t _dststride, | |
1457 | const uint8_t *_src, ptrdiff_t _srcstride, const int16_t *src2, | ||
1458 | int height, int denom, int wx0, int wx1, | ||
1459 | int ox0, int ox1, intptr_t mx, intptr_t my, int width) | ||
1460 | { | ||
1461 | int x, y; | ||
1462 | 166916 | const pixel *src = (const pixel *)_src; | |
1463 | 166916 | ptrdiff_t srcstride = _srcstride / sizeof(pixel); | |
1464 | 166916 | pixel *dst = (pixel *)_dst; | |
1465 | 166916 | ptrdiff_t dststride = _dststride / sizeof(pixel); | |
1466 | 166916 | const int8_t *filter = ff_hevc_epel_filters[mx - 1]; | |
1467 | int16_t tmp_array[(MAX_PB_SIZE + EPEL_EXTRA) * MAX_PB_SIZE]; | ||
1468 | 166916 | int16_t *tmp = tmp_array; | |
1469 | 166916 | int shift = 14 + 1 - BIT_DEPTH; | |
1470 | 166916 | int log2Wd = denom + shift - 1; | |
1471 | |||
1472 | 166916 | src -= EPEL_EXTRA_BEFORE * srcstride; | |
1473 | |||
1474 |
2/2✓ Branch 0 taken 1288910 times.
✓ Branch 1 taken 83458 times.
|
2744736 | for (y = 0; y < height + EPEL_EXTRA; y++) { |
1475 |
2/2✓ Branch 0 taken 21782220 times.
✓ Branch 1 taken 1288910 times.
|
46142260 | for (x = 0; x < width; x++) |
1476 | 43564440 | tmp[x] = EPEL_FILTER(src, 1) >> (BIT_DEPTH - 8); | |
1477 | 2577820 | src += srcstride; | |
1478 | 2577820 | tmp += MAX_PB_SIZE; | |
1479 | } | ||
1480 | |||
1481 | 166916 | tmp = tmp_array + EPEL_EXTRA_BEFORE * MAX_PB_SIZE; | |
1482 | 166916 | filter = ff_hevc_epel_filters[my - 1]; | |
1483 | |||
1484 | 166916 | ox0 = ox0 * (1 << (BIT_DEPTH - 8)); | |
1485 | 166916 | ox1 = ox1 * (1 << (BIT_DEPTH - 8)); | |
1486 |
2/2✓ Branch 0 taken 1038536 times.
✓ Branch 1 taken 83458 times.
|
2243988 | for (y = 0; y < height; y++) { |
1487 |
2/2✓ Branch 0 taken 18780552 times.
✓ Branch 1 taken 1038536 times.
|
39638176 | for (x = 0; x < width; x++) |
1488 | 37561104 | dst[x] = av_clip_pixel(((EPEL_FILTER(tmp, MAX_PB_SIZE) >> 6) * wx1 + src2[x] * wx0 + | |
1489 | ((ox0 + ox1 + 1) * (1 << log2Wd))) >> (log2Wd + 1)); | ||
1490 | 2077072 | tmp += MAX_PB_SIZE; | |
1491 | 2077072 | dst += dststride; | |
1492 | 2077072 | src2 += MAX_PB_SIZE; | |
1493 | } | ||
1494 | 166916 | } | |
1495 | |||
1496 | // line zero | ||
1497 | #define P3 pix[-4 * xstride] | ||
1498 | #define P2 pix[-3 * xstride] | ||
1499 | #define P1 pix[-2 * xstride] | ||
1500 | #define P0 pix[-1 * xstride] | ||
1501 | #define Q0 pix[0 * xstride] | ||
1502 | #define Q1 pix[1 * xstride] | ||
1503 | #define Q2 pix[2 * xstride] | ||
1504 | #define Q3 pix[3 * xstride] | ||
1505 | |||
1506 | // line three. used only for deblocking decision | ||
1507 | #define TP3 pix[-4 * xstride + 3 * ystride] | ||
1508 | #define TP2 pix[-3 * xstride + 3 * ystride] | ||
1509 | #define TP1 pix[-2 * xstride + 3 * ystride] | ||
1510 | #define TP0 pix[-1 * xstride + 3 * ystride] | ||
1511 | #define TQ0 pix[0 * xstride + 3 * ystride] | ||
1512 | #define TQ1 pix[1 * xstride + 3 * ystride] | ||
1513 | #define TQ2 pix[2 * xstride + 3 * ystride] | ||
1514 | #define TQ3 pix[3 * xstride + 3 * ystride] | ||
1515 | |||
1516 | 69349318 | static void FUNC(hevc_loop_filter_luma)(uint8_t *_pix, | |
1517 | ptrdiff_t _xstride, ptrdiff_t _ystride, | ||
1518 | int beta, const int *_tc, | ||
1519 | const uint8_t *_no_p, const uint8_t *_no_q) | ||
1520 | { | ||
1521 | int d, j; | ||
1522 | 69349318 | pixel *pix = (pixel *)_pix; | |
1523 | 69349318 | ptrdiff_t xstride = _xstride / sizeof(pixel); | |
1524 | 69349318 | ptrdiff_t ystride = _ystride / sizeof(pixel); | |
1525 | |||
1526 | 69349318 | beta <<= BIT_DEPTH - 8; | |
1527 | |||
1528 |
2/2✓ Branch 0 taken 69349318 times.
✓ Branch 1 taken 34674659 times.
|
208047954 | for (j = 0; j < 2; j++) { |
1529 | 138698636 | const int dp0 = abs(P2 - 2 * P1 + P0); | |
1530 | 138698636 | const int dq0 = abs(Q2 - 2 * Q1 + Q0); | |
1531 | 138698636 | const int dp3 = abs(TP2 - 2 * TP1 + TP0); | |
1532 | 138698636 | const int dq3 = abs(TQ2 - 2 * TQ1 + TQ0); | |
1533 | 138698636 | const int d0 = dp0 + dq0; | |
1534 | 138698636 | const int d3 = dp3 + dq3; | |
1535 | 138698636 | const int tc = _tc[j] << (BIT_DEPTH - 8); | |
1536 | 138698636 | const int no_p = _no_p[j]; | |
1537 | 138698636 | const int no_q = _no_q[j]; | |
1538 | |||
1539 |
2/2✓ Branch 0 taken 17977241 times.
✓ Branch 1 taken 51372077 times.
|
138698636 | if (d0 + d3 >= beta) { |
1540 | 35954482 | pix += 4 * ystride; | |
1541 | 35954482 | continue; | |
1542 | } else { | ||
1543 | 102744154 | const int beta_3 = beta >> 3; | |
1544 | 102744154 | const int beta_2 = beta >> 2; | |
1545 | 102744154 | const int tc25 = ((tc * 5 + 1) >> 1); | |
1546 | |||
1547 |
4/4✓ Branch 0 taken 14506670 times.
✓ Branch 1 taken 36865407 times.
✓ Branch 2 taken 14057788 times.
✓ Branch 3 taken 448882 times.
|
102744154 | if (abs(P3 - P0) + abs(Q3 - Q0) < beta_3 && abs(P0 - Q0) < tc25 && |
1548 |
4/4✓ Branch 0 taken 10521946 times.
✓ Branch 1 taken 3535842 times.
✓ Branch 2 taken 10430063 times.
✓ Branch 3 taken 91883 times.
|
28115576 | abs(TP3 - TP0) + abs(TQ3 - TQ0) < beta_3 && abs(TP0 - TQ0) < tc25 && |
1549 |
4/4✓ Branch 0 taken 10072655 times.
✓ Branch 1 taken 357408 times.
✓ Branch 2 taken 9858123 times.
✓ Branch 3 taken 214532 times.
|
40576372 | (d0 << 1) < beta_2 && (d3 << 1) < beta_2) { |
1550 | // strong filtering | ||
1551 | 19716246 | const int tc2 = tc << 1; | |
1552 |
2/2✓ Branch 0 taken 39432492 times.
✓ Branch 1 taken 9858123 times.
|
98581230 | for (d = 0; d < 4; d++) { |
1553 | 78864984 | const int p3 = P3; | |
1554 | 78864984 | const int p2 = P2; | |
1555 | 78864984 | const int p1 = P1; | |
1556 | 78864984 | const int p0 = P0; | |
1557 | 78864984 | const int q0 = Q0; | |
1558 | 78864984 | const int q1 = Q1; | |
1559 | 78864984 | const int q2 = Q2; | |
1560 | 78864984 | const int q3 = Q3; | |
1561 |
2/2✓ Branch 0 taken 39426948 times.
✓ Branch 1 taken 5544 times.
|
78864984 | if (!no_p) { |
1562 | 78853896 | P0 = p0 + av_clip(((p2 + 2 * p1 + 2 * p0 + 2 * q0 + q1 + 4) >> 3) - p0, -tc2, tc2); | |
1563 | 78853896 | P1 = p1 + av_clip(((p2 + p1 + p0 + q0 + 2) >> 2) - p1, -tc2, tc2); | |
1564 | 78853896 | P2 = p2 + av_clip(((2 * p3 + 3 * p2 + p1 + p0 + q0 + 4) >> 3) - p2, -tc2, tc2); | |
1565 | } | ||
1566 |
2/2✓ Branch 0 taken 39426820 times.
✓ Branch 1 taken 5672 times.
|
78864984 | if (!no_q) { |
1567 | 78853640 | Q0 = q0 + av_clip(((p1 + 2 * p0 + 2 * q0 + 2 * q1 + q2 + 4) >> 3) - q0, -tc2, tc2); | |
1568 | 78853640 | Q1 = q1 + av_clip(((p0 + q0 + q1 + q2 + 2) >> 2) - q1, -tc2, tc2); | |
1569 | 78853640 | Q2 = q2 + av_clip(((2 * q3 + 3 * q2 + q1 + q0 + p0 + 4) >> 3) - q2, -tc2, tc2); | |
1570 | } | ||
1571 | 78864984 | pix += ystride; | |
1572 | } | ||
1573 | } else { // normal filtering | ||
1574 | 83027908 | int nd_p = 1; | |
1575 | 83027908 | int nd_q = 1; | |
1576 | 83027908 | const int tc_2 = tc >> 1; | |
1577 |
2/2✓ Branch 0 taken 29180467 times.
✓ Branch 1 taken 12333487 times.
|
83027908 | if (dp0 + dp3 < ((beta + (beta >> 1)) >> 3)) |
1578 | 58360934 | nd_p = 2; | |
1579 |
2/2✓ Branch 0 taken 28407550 times.
✓ Branch 1 taken 13106404 times.
|
83027908 | if (dq0 + dq3 < ((beta + (beta >> 1)) >> 3)) |
1580 | 56815100 | nd_q = 2; | |
1581 | |||
1582 |
2/2✓ Branch 0 taken 166055816 times.
✓ Branch 1 taken 41513954 times.
|
415139540 | for (d = 0; d < 4; d++) { |
1583 | 332111632 | const int p2 = P2; | |
1584 | 332111632 | const int p1 = P1; | |
1585 | 332111632 | const int p0 = P0; | |
1586 | 332111632 | const int q0 = Q0; | |
1587 | 332111632 | const int q1 = Q1; | |
1588 | 332111632 | const int q2 = Q2; | |
1589 | 332111632 | int delta0 = (9 * (q0 - p0) - 3 * (q1 - p1) + 8) >> 4; | |
1590 |
2/2✓ Branch 0 taken 164217023 times.
✓ Branch 1 taken 1838793 times.
|
332111632 | if (abs(delta0) < 10 * tc) { |
1591 | 328434046 | delta0 = av_clip(delta0, -tc, tc); | |
1592 |
2/2✓ Branch 0 taken 163901706 times.
✓ Branch 1 taken 315317 times.
|
328434046 | if (!no_p) |
1593 | 327803412 | P0 = av_clip_pixel(p0 + delta0); | |
1594 |
2/2✓ Branch 0 taken 163899775 times.
✓ Branch 1 taken 317248 times.
|
328434046 | if (!no_q) |
1595 | 327799550 | Q0 = av_clip_pixel(q0 - delta0); | |
1596 |
4/4✓ Branch 0 taken 163901706 times.
✓ Branch 1 taken 315317 times.
✓ Branch 2 taken 115525206 times.
✓ Branch 3 taken 48376500 times.
|
328434046 | if (!no_p && nd_p > 1) { |
1597 | 231050412 | const int deltap1 = av_clip((((p2 + p0 + 1) >> 1) - p1 + delta0) >> 1, -tc_2, tc_2); | |
1598 | 231050412 | P1 = av_clip_pixel(p1 + deltap1); | |
1599 | } | ||
1600 |
4/4✓ Branch 0 taken 163899775 times.
✓ Branch 1 taken 317248 times.
✓ Branch 2 taken 112455172 times.
✓ Branch 3 taken 51444603 times.
|
328434046 | if (!no_q && nd_q > 1) { |
1601 | 224910344 | const int deltaq1 = av_clip((((q2 + q0 + 1) >> 1) - q1 - delta0) >> 1, -tc_2, tc_2); | |
1602 | 224910344 | Q1 = av_clip_pixel(q1 + deltaq1); | |
1603 | } | ||
1604 | } | ||
1605 | 332111632 | pix += ystride; | |
1606 | } | ||
1607 | } | ||
1608 | } | ||
1609 | } | ||
1610 | 69349318 | } | |
1611 | |||
1612 | 21771428 | static void FUNC(hevc_loop_filter_chroma)(uint8_t *_pix, ptrdiff_t _xstride, | |
1613 | ptrdiff_t _ystride, const int *_tc, | ||
1614 | const uint8_t *_no_p, const uint8_t *_no_q) | ||
1615 | { | ||
1616 | int d, j, no_p, no_q; | ||
1617 | 21771428 | pixel *pix = (pixel *)_pix; | |
1618 | 21771428 | ptrdiff_t xstride = _xstride / sizeof(pixel); | |
1619 | 21771428 | ptrdiff_t ystride = _ystride / sizeof(pixel); | |
1620 | |||
1621 |
2/2✓ Branch 0 taken 21771428 times.
✓ Branch 1 taken 10885714 times.
|
65314284 | for (j = 0; j < 2; j++) { |
1622 | 43542856 | const int tc = _tc[j] << (BIT_DEPTH - 8); | |
1623 |
2/2✓ Branch 0 taken 2274659 times.
✓ Branch 1 taken 19496769 times.
|
43542856 | if (tc <= 0) { |
1624 | 4549318 | pix += 4 * ystride; | |
1625 | 4549318 | continue; | |
1626 | } | ||
1627 | 38993538 | no_p = _no_p[j]; | |
1628 | 38993538 | no_q = _no_q[j]; | |
1629 | |||
1630 |
2/2✓ Branch 0 taken 77987076 times.
✓ Branch 1 taken 19496769 times.
|
194967690 | for (d = 0; d < 4; d++) { |
1631 | int delta0; | ||
1632 | 155974152 | const int p1 = P1; | |
1633 | 155974152 | const int p0 = P0; | |
1634 | 155974152 | const int q0 = Q0; | |
1635 | 155974152 | const int q1 = Q1; | |
1636 | 155974152 | delta0 = av_clip((((q0 - p0) * 4) + p1 - q1 + 4) >> 3, -tc, tc); | |
1637 |
2/2✓ Branch 0 taken 77724484 times.
✓ Branch 1 taken 262592 times.
|
155974152 | if (!no_p) |
1638 | 155448968 | P0 = av_clip_pixel(p0 + delta0); | |
1639 |
2/2✓ Branch 0 taken 77720628 times.
✓ Branch 1 taken 266448 times.
|
155974152 | if (!no_q) |
1640 | 155441256 | Q0 = av_clip_pixel(q0 - delta0); | |
1641 | 155974152 | pix += ystride; | |
1642 | } | ||
1643 | } | ||
1644 | 21771428 | } | |
1645 | |||
1646 | 10600800 | static void FUNC(hevc_h_loop_filter_chroma)(uint8_t *pix, ptrdiff_t stride, | |
1647 | const int32_t *tc, const uint8_t *no_p, | ||
1648 | const uint8_t *no_q) | ||
1649 | { | ||
1650 | 10600800 | FUNC(hevc_loop_filter_chroma)(pix, stride, sizeof(pixel), tc, no_p, no_q); | |
1651 | 10600800 | } | |
1652 | |||
1653 | 11170628 | static void FUNC(hevc_v_loop_filter_chroma)(uint8_t *pix, ptrdiff_t stride, | |
1654 | const int32_t *tc, const uint8_t *no_p, | ||
1655 | const uint8_t *no_q) | ||
1656 | { | ||
1657 | 11170628 | FUNC(hevc_loop_filter_chroma)(pix, sizeof(pixel), stride, tc, no_p, no_q); | |
1658 | 11170628 | } | |
1659 | |||
1660 | 35009204 | static void FUNC(hevc_h_loop_filter_luma)(uint8_t *pix, ptrdiff_t stride, | |
1661 | int beta, const int32_t *tc, const uint8_t *no_p, | ||
1662 | const uint8_t *no_q) | ||
1663 | { | ||
1664 | 35009204 | FUNC(hevc_loop_filter_luma)(pix, stride, sizeof(pixel), | |
1665 | beta, tc, no_p, no_q); | ||
1666 | 35009204 | } | |
1667 | |||
1668 | 34340114 | static void FUNC(hevc_v_loop_filter_luma)(uint8_t *pix, ptrdiff_t stride, | |
1669 | int beta, const int32_t *tc, const uint8_t *no_p, | ||
1670 | const uint8_t *no_q) | ||
1671 | { | ||
1672 | 34340114 | FUNC(hevc_loop_filter_luma)(pix, sizeof(pixel), stride, | |
1673 | beta, tc, no_p, no_q); | ||
1674 | 34340114 | } | |
1675 | |||
1676 | #undef P3 | ||
1677 | #undef P2 | ||
1678 | #undef P1 | ||
1679 | #undef P0 | ||
1680 | #undef Q0 | ||
1681 | #undef Q1 | ||
1682 | #undef Q2 | ||
1683 | #undef Q3 | ||
1684 | |||
1685 | #undef TP3 | ||
1686 | #undef TP2 | ||
1687 | #undef TP1 | ||
1688 | #undef TP0 | ||
1689 | #undef TQ0 | ||
1690 | #undef TQ1 | ||
1691 | #undef TQ2 | ||
1692 | #undef TQ3 | ||
1693 |